Python Cookbook, 2nd Edition

f74f

贡献于2015-01-21

字数:0 关键词: Python开发 Python

• Table of Contents • Index • Reviews • Reader Reviews • Errata • Academic Python Cookbook, 2nd Edition By David Ascher, Alex Martelli, Anna Ravenscroft Publisher: O'Reilly Pub Date: March 2005 ISBN: 0-596-00797-3 Pages: 844 Like its predecessor, the new edition offers a collection of solutions to problems that Python programmers face everyday. Updated for Python 2.4, it now includes over 200 recipes that range from simple tasks, such as working with dictionaries and list comprehensions, to complex tasks, such as monitoring a network and building a templating system. • Table of Contents • Index • Reviews • Reader Reviews • Errata • Academic Python Cookbook, 2nd Edition By David Ascher, Alex Martelli, Anna Ravenscroft Publisher: O'Reilly Pub Date: March 2005 ISBN: 0-596-00797-3 Pages: 844 Copyright Preface The Design of the Book The Implementation of the Book Using the Code from This Book Audience Organization Further Reading Conventions Used in This Book How to Contact Us Safari® Enabled Acknowledgments Chapter 1. Text Introduction Recipe 1.1. Processing a String One Character at a Time Recipe 1.2. Converting Between Characters and Numeric Codes Recipe 1.3. Testing Whether an Object Is String-like Recipe 1.4. Aligning Strings Recipe 1.5. Trimming Space from the Ends of a String Recipe 1.6. Combining Strings Recipe 1.7. Reversing a String by Words or Characters Recipe 1.8. Checking Whether a String Contains a Set of Characters Recipe 1.9. Simplifying Usage of Strings' translate Method Recipe 1.10. Filtering a String for a Set of Characters Recipe 1.11. Checking Whether a String Is Text or Binary Recipe 1.12. Controlling Case Recipe 1.13. Accessing Substrings Recipe 1.14. Changing the Indentation of a Multiline String Recipe 1.15. Expanding and Compressing Tabs Recipe 1.16. Interpolating Variables in a String Recipe 1.17. Interpolating Variables in a Stringin Python 2.4 Recipe 1.18. Replacing Multiple Patterns in a Single Pass Recipe 1.19. Checking a String for Any of Multiple Endings Recipe 1.20. Handling International Text with Unicode Recipe 1.21. Converting Between Unicode and Plain Strings Recipe 1.22. Printing Unicode Charactersto Standard Output Recipe 1.23. Encoding Unicode Data for XML and HTML Recipe 1.24. Making Some Strings Case-Insensitive Recipe 1.25. Converting HTML Documents to Texton a Unix Terminal Chapter 2. Files Introduction Recipe 2.1. Reading from a File Recipe 2.2. Writing to a File Recipe 2.3. Searching and Replacing Text in a File Recipe 2.4. Reading a Specific Line from a File Recipe 2.5. Counting Lines in a File Recipe 2.6. Processing Every Word in a File Recipe 2.7. Using Random-Access Input/Output Recipe 2.8. Updating a Random-Access File Recipe 2.9. Reading Data from zip Files Recipe 2.10. Handling a zip File Inside a String Recipe 2.11. Archiving a Tree of Files into a Compressed tar File Recipe 2.12. Sending Binary Data to Standard Output Under Windows Recipe 2.13. Using a C++-like iostream Syntax Recipe 2.14. Rewinding an Input File to the Beginning Recipe 2.15. Adapting a File-like Object to a True File Object Recipe 2.16. Walking Directory Trees Recipe 2.17. Swapping One File Extension for Another Throughout a Directory Tree Recipe 2.18. Finding a File Given a Search Path Recipe 2.19. Finding Files Given a Search Path and a Pattern Recipe 2.20. Finding a File on the Python Search Path Recipe 2.21. Dynamically Changing the PythonSearch Path Recipe 2.22. Computing the Relative Path from One Directory to Another Recipe 2.23. Reading an Unbuffered Character in a Cross-Platform Way Recipe 2.24. Counting Pages of PDF Documents on Mac OS X Recipe 2.25. Changing File Attributes on Windows Recipe 2.26. Extracting Text from OpenOffice.org Documents Recipe 2.27. Extracting Text from Microsoft Word Documents Recipe 2.28. File Locking Using a Cross-Platform API Recipe 2.29. Versioning Filenames Recipe 2.30. Calculating CRC-64 Cyclic Redundancy Checks Chapter 3. Time and Money Introduction Recipe 3.1. Calculating Yesterday and Tomorrow Recipe 3.2. Finding Last Friday Recipe 3.3. Calculating Time Periods in a Date Range Recipe 3.4. Summing Durations of Songs Recipe 3.5. Calculating the Number of Weekdays Between Two Dates Recipe 3.6. Looking up Holidays Automatically Recipe 3.7. Fuzzy Parsing of Dates Recipe 3.8. Checking Whether Daylight Saving Time Is Currently in Effect Recipe 3.9. Converting Time Zones Recipe 3.10. Running a Command Repeatedly Recipe 3.11. Scheduling Commands Recipe 3.12. Doing Decimal Arithmetic Recipe 3.13. Formatting Decimals as Currency Recipe 3.14. Using Python as a Simple Adding Machine Recipe 3.15. Checking a Credit Card Checksum Recipe 3.16. Watching Foreign Exchange Rates Chapter 4. Python Shortcuts Introduction Recipe 4.1. Copying an Object Recipe 4.2. Constructing Lists with List Comprehensions Recipe 4.3. Returning an Element of a List If It Exists Recipe 4.4. Looping over Items and Their Indices in a Sequence Recipe 4.5. Creating Lists of Lists Without Sharing References Recipe 4.6. Flattening a Nested Sequence Recipe 4.7. Removing or Reordering Columnsin a List of Rows Recipe 4.8. Transposing Two-Dimensional Arrays Recipe 4.9. Getting a Value from a Dictionary Recipe 4.10. Adding an Entry to a Dictionary Recipe 4.11. Building a Dictionary Without Excessive Quoting Recipe 4.12. Building a Dict from a List of Alternating Keys and Values Recipe 4.13. Extracting a Subset of a Dictionary Recipe 4.14. Inverting a Dictionary Recipe 4.15. Associating Multiple Values with Each Key in a Dictionary Recipe 4.16. Using a Dictionary to Dispatch Methods or Functions Recipe 4.17. Finding Unions and Intersections of Dictionaries Recipe 4.18. Collecting a Bunch of Named Items Recipe 4.19. Assigning and Testing with One Statement Recipe 4.20. Using printf in Python Recipe 4.21. Randomly Picking Items with Given Probabilities Recipe 4.22. Handling Exceptions Within an Expression Recipe 4.23. Ensuring a Name Is Defined in a Given Module Chapter 5. Searching and Sorting Introduction Recipe 5.1. Sorting a Dictionary Recipe 5.2. Sorting a List of Strings Case-Insensitively Recipe 5.3. Sorting a List of Objects by an Attribute of the Objects Recipe 5.4. Sorting Keys or Indices Basedon the Corresponding Values Recipe 5.5. Sorting Strings with Embedded Numbers Recipe 5.6. Processing All of a List's Items in Random Order Recipe 5.7. Keeping a Sequence Ordered as Items Are Added Recipe 5.8. Getting the First Few Smallest Items of a Sequence Recipe 5.9. Looking for Items in a Sorted Sequence Recipe 5.10. Selecting the nth Smallest Element of a Sequence Recipe 5.11. Showing off quicksort in Three Lines Recipe 5.12. Performing Frequent Membership Tests on a Sequence Recipe 5.13. Finding Subsequences Recipe 5.14. Enriching the Dictionary Type with Ratings Functionality Recipe 5.15. Sorting Names and Separating Them by Initials Chapter 6. Object-Oriented Programming Introduction Recipe 6.1. Converting Among Temperature Scales Recipe 6.2. Defining Constants Recipe 6.3. Restricting Attribute Setting Recipe 6.4. Chaining Dictionary Lookups Recipe 6.5. Delegating Automatically as an Alternative to Inheritance Recipe 6.6. Delegating Special Methods in Proxies Recipe 6.7. Implementing Tuples with Named Items Recipe 6.8. Avoiding Boilerplate Accessors for Properties Recipe 6.9. Making a Fast Copy of an Object Recipe 6.10. Keeping References to Bound Methods Without Inhibiting Garbage Collection Recipe 6.11. Implementing a Ring Buffer Recipe 6.12. Checking an Instance for Any State Changes Recipe 6.13. Checking Whether an Object Has Necessary Attributes Recipe 6.14. Implementing the State Design Pattern Recipe 6.15. Implementing the "Singleton" Design Pattern Recipe 6.16. Avoiding the "Singleton" Design Pattern with the Borg Idiom Recipe 6.17. Implementing the Null Object Design Pattern Recipe 6.18. Automatically Initializing Instance Variables from _ _init_ _ Arguments Recipe 6.19. Calling a Superclass _ _init_ _ Method If It Exists Recipe 6.20. Using Cooperative Supercalls Concisely and Safely Chapter 7. Persistence and Databases Introduction Recipe 7.1. Serializing Data Using the marshal Module Recipe 7.2. Serializing Data Using the pickle and cPickle Modules Recipe 7.3. Using Compression with Pickling Recipe 7.4. Using the cPickle Module on Classes and Instances Recipe 7.5. Holding Bound Methods in a Picklable Way Recipe 7.6. Pickling Code Objects Recipe 7.7. Mutating Objects with shelve Recipe 7.8. Using the Berkeley DB Database Recipe 7.9. Accesssing a MySQL Database Recipe 7.10. Storing a BLOB in a MySQL Database Recipe 7.11. Storing a BLOB in a PostgreSQL Database Recipe 7.12. Storing a BLOB in a SQLite Database Recipe 7.13. Generating a Dictionary Mapping Field Names to Column Numbers Recipe 7.14. Using dtuple for Flexible Accessto Query Results Recipe 7.15. Pretty-Printing the Contents of Database Cursors Recipe 7.16. Using a Single Parameter-Passing Style Across Various DB API Modules Recipe 7.17. Using Microsoft Jet via ADO Recipe 7.18. Accessing a JDBC Database from a Jython Servlet Recipe 7.19. Using ODBC to Get Excel Data with Jython Chapter 8. Debugging and Testing Introduction Recipe 8.1. Disabling Execution of Some Conditionals and Loops Recipe 8.2. Measuring Memory Usage on Linux Recipe 8.3. Debugging the Garbage-Collection Process Recipe 8.4. Trapping and Recording Exceptions Recipe 8.5. Tracing Expressions and Comments in Debug Mode Recipe 8.6. Getting More Information from Tracebacks Recipe 8.7. Starting the Debugger Automatically After an Uncaught Exception Recipe 8.8. Running Unit Tests Most Simply Recipe 8.9. Running Unit Tests Automatically Recipe 8.10. Using doctest with unittest in Python 2.4 Recipe 8.11. Checking Values Against Intervals in Unit Testing Chapter 9. Processes, Threads, and Synchronization Introduction Recipe 9.1. Synchronizing All Methods in an Object Recipe 9.2. Terminating a Thread Recipe 9.3. Using a Queue.Queue as a Priority Queue Recipe 9.4. Working with a Thread Pool Recipe 9.5. Executing a Function in Parallel on Multiple Argument Sets Recipe 9.6. Coordinating Threads by Simple Message Passing Recipe 9.7. Storing Per-Thread Information Recipe 9.8. Multitasking Cooperatively Without Threads Recipe 9.9. Determining Whether Another Instanceof a Script Is Already Running in Windows Recipe 9.10. Processing Windows Messages Using MsgWaitForMultipleObjects Recipe 9.11. Driving an External Process with popen Recipe 9.12. Capturing the Output and Error Streams from a Unix Shell Command Recipe 9.13. Forking a Daemon Process on Unix Chapter 10. System Administration Introduction Recipe 10.1. Generating Random Passwords Recipe 10.2. Generating Easily Remembered Somewhat-Random Passwords Recipe 10.3. Authenticating Users by Means of a POP Server Recipe 10.4. Calculating Apache Hits per IP Address Recipe 10.5. Calculating the Rate of Client Cache Hits on Apache Recipe 10.6. Spawning an Editor from a Script Recipe 10.7. Backing Up Files Recipe 10.8. Selectively Copying a Mailbox File Recipe 10.9. Building a Whitelist of Email Addresses From a Mailbox Recipe 10.10. Blocking Duplicate Mails Recipe 10.11. Checking Your Windows Sound System Recipe 10.12. Registering or Unregistering a DLL on Windows Recipe 10.13. Checking and Modifying the Set of Tasks Windows Automatically Runs at Login Recipe 10.14. Creating a Share on Windows Recipe 10.15. Connecting to an Already Running Instance of Internet Explorer Recipe 10.16. Reading Microsoft Outlook Contacts Recipe 10.17. Gathering Detailed System Informationon Mac OS X Chapter 11. User Interfaces Introduction Recipe 11.1. Showing a Progress Indicator on a Text Console Recipe 11.2. Avoiding lambda in Writing Callback Functions Recipe 11.3. Using Default Values and Bounds with tkSimpleDialog Functions Recipe 11.4. Adding Drag and Drop Reordering to a Tkinter Listbox Recipe 11.5. Entering Accented Characters in Tkinter Widgets Recipe 11.6. Embedding Inline GIFs Using Tkinter Recipe 11.7. Converting Among Image Formats Recipe 11.8. Implementing a Stopwatch in Tkinter Recipe 11.9. Combining GUIs and Asynchronous I/Owith Threads Recipe 11.10. Using IDLE's Tree Widget in Tkinter Recipe 11.11. Supporting Multiple Values per Row in a Tkinter Listbox Recipe 11.12. Copying Geometry Methods and Options Between Tkinter Widgets Recipe 11.13. Implementing a Tabbed Notebook for Tkinter Recipe 11.14. Using a wxPython Notebook with Panels Recipe 11.15. Implementing an ImageJ Plug-in in Jython Recipe 11.16. Viewing an Image from a URL with Swing and Jython Recipe 11.17. Getting User Input on Mac OS Recipe 11.18. Building a Python Cocoa GUI Programmatically Recipe 11.19. Implementing Fade-in Windows with IronPython Chapter 12. Processing XML Introduction Recipe 12.1. Checking XML Well-Formedness Recipe 12.2. Counting Tags in a Document Recipe 12.3. Extracting Text from an XML Document Recipe 12.4. Autodetecting XML Encoding Recipe 12.5. Converting an XML Document into a Tree of Python Objects Recipe 12.6. Removing Whitespace-only Text Nodes from an XML DOM Node's Subtree Recipe 12.7. Parsing Microsoft Excel's XML Recipe 12.8. Validating XML Documents Recipe 12.9. Filtering Elements and Attributes Belonging to a Given Namespace Recipe 12.10. Merging Continuous Text Events with a SAX Filter Recipe 12.11. Using MSHTML to Parse XML or HTML Chapter 13. Network Programming Introduction Recipe 13.1. Passing Messages with Socket Datagrams Recipe 13.2. Grabbing a Document from the Web Recipe 13.3. Filtering a List of FTP Sites Recipe 13.4. Getting Time from a Server via the SNTP Protocol Recipe 13.5. Sending HTML Mail Recipe 13.6. Bundling Files in a MIME Message Recipe 13.7. Unpacking a Multipart MIME Message Recipe 13.8. Removing Attachments from an Email Message Recipe 13.9. Fixing Messages Parsed by Python 2.4 email.FeedParser Recipe 13.10. Inspecting a POP3 Mailbox Interactively Recipe 13.11. Detecting Inactive Computers Recipe 13.12. Monitoring a Network with HTTP Recipe 13.13. Forwarding and Redirecting Network Ports Recipe 13.14. Tunneling SSL Through a Proxy Recipe 13.15. Implementing the Dynamic IP Protocol Recipe 13.16. Connecting to IRC and Logging Messages to Disk Recipe 13.17. Accessing LDAP Servers Chapter 14. Web Programming Introduction Recipe 14.1. Testing Whether CGI Is Working Recipe 14.2. Handling URLs Within a CGI Script Recipe 14.3. Uploading Files with CGI Recipe 14.4. Checking for a Web Page's Existence Recipe 14.5. Checking Content Type via HTTP Recipe 14.6. Resuming the HTTP Download of a File Recipe 14.7. Handling Cookies While Fetching Web Pages Recipe 14.8. Authenticating with a Proxy for HTTPS Navigation Recipe 14.9. Running a Servlet with Jython Recipe 14.10. Finding an Internet Explorer Cookie Recipe 14.11. Generating OPML Files Recipe 14.12. Aggregating RSS Feeds Recipe 14.13. Turning Data into Web Pages Through Templates Recipe 14.14. Rendering Arbitrary Objects with Nevow Chapter 15. Distributed Programming Introduction Recipe 15.1. Making an XML-RPC Method Call Recipe 15.2. Serving XML-RPC Requests Recipe 15.3. Using XML-RPC with Medusa Recipe 15.4. Enabling an XML-RPC Server to Be Terminated Remotely Recipe 15.5. Implementing SimpleXMLRPCServer Niceties Recipe 15.6. Giving an XML-RPC Server a wxPython GUI Recipe 15.7. Using Twisted Perspective Broker Recipe 15.8. Implementing a CORBA Server and Client Recipe 15.9. Performing Remote Logins Using telnetlib Recipe 15.10. Performing Remote Logins with SSH Recipe 15.11. Authenticating an SSL Client over HTTPS Chapter 16. Programs About Programs Introduction Recipe 16.1. Verifying Whether a String Represents a Valid Number Recipe 16.2. Importing a Dynamically Generated Module Recipe 16.3. Importing from a Module Whose Name Is Determined at Runtime Recipe 16.4. Associating Parameters with a Function (Currying) Recipe 16.5. Composing Functions Recipe 16.6. Colorizing Python Source Using the Built-in Tokenizer Recipe 16.7. Merging and Splitting Tokens Recipe 16.8. Checking Whether a String Has Balanced Parentheses Recipe 16.9. Simulating Enumerations in Python Recipe 16.10. Referring to a List Comprehension While Building It Recipe 16.11. Automating the py2exe Compilation of Scripts into Windows Executables Recipe 16.12. Binding Main Script and Modules into One Executable on Unix Chapter 17. Extending and Embedding Introduction Recipe 17.1. Implementing a Simple Extension Type Recipe 17.2. Implementing a Simple Extension Type with Pyrex Recipe 17.3. Exposing a C++ Library to Python Recipe 17.4. Calling Functions from a Windows DLL Recipe 17.5. Using SWIG-Generated Modules in a Multithreaded Environment Recipe 17.6. Translating a Python Sequence into a C Array with the PySequence_Fast Protocol Recipe 17.7. Accessing a Python Sequence Item-by-Item with the Iterator Protocol Recipe 17.8. Returning None from a Python-Callable C Function Recipe 17.9. Debugging Dynamically Loaded C Extensions with gdb Recipe 17.10. Debugging Memory Problems Chapter 18. Algorithms Introduction Recipe 18.1. Removing Duplicates from a Sequence Recipe 18.2. Removing Duplicates from a Sequence While Maintaining Sequence Order Recipe 18.3. Generating Random Samples with Replacement Recipe 18.4. Generating Random Samples Without Replacement Recipe 18.5. Memoizing (Caching) the Return Values of Functions Recipe 18.6. Implementing a FIFO Container Recipe 18.7. Caching Objects with a FIFO Pruning Strategy Recipe 18.8. Implementing a Bag (Multiset) Collection Type Recipe 18.9. Simulating the Ternary Operator in Python Recipe 18.10. Computing Prime Numbers Recipe 18.11. Formatting Integers as Binary Strings Recipe 18.12. Formatting Integers as Strings in Arbitrary Bases Recipe 18.13. Converting Numbers to Rationals via Farey Fractions Recipe 18.14. Doing Arithmetic with Error Propagation Recipe 18.15. Summing Numbers with Maximal Accuracy Recipe 18.16. Simulating Floating Point Recipe 18.17. Computing the Convex Hulls and Diameters of 2D Point Sets Chapter 19. Iterators and Generators Introduction Recipe 19.1. Writing a range-like Function with Float Increments Recipe 19.2. Building a List from Any Iterable Recipe 19.3. Generating the Fibonacci Sequence Recipe 19.4. Unpacking a Few Items in a Multiple Assignment Recipe 19.5. Automatically Unpacking the Needed Number of Items Recipe 19.6. Dividing an Iterable into n Slices of Stride n Recipe 19.7. Looping on a Sequence by Overlapping Windows Recipe 19.8. Looping Through Multiple Iterables in Parallel Recipe 19.9. Looping Through the Cross-Product of Multiple Iterables Recipe 19.10. Reading a Text File by Paragraphs Recipe 19.11. Reading Lines with Continuation Characters Recipe 19.12. Iterating on a Stream of Data Blocks as a Stream of Lines Recipe 19.13. Fetching Large Record Sets from a Database with a Generator Recipe 19.14. Merging Sorted Sequences Recipe 19.15. Generating Permutations, Combinations, and Selections Recipe 19.16. Generating the Partitions of an Integer Recipe 19.17. Duplicating an Iterator Recipe 19.18. Looking Ahead into an Iterator Recipe 19.19. Simplifying Queue-Consumer Threads Recipe 19.20. Running an Iterator in Another Thread Recipe 19.21. Computing a Summary Report with itertools.groupby Chapter 20. Descriptors, Decorators,and Metaclasses Introduction Recipe 20.1. Getting Fresh Default Values at Each Function Call Recipe 20.2. Coding Properties as Nested Functions Recipe 20.3. Aliasing Attribute Values Recipe 20.4. Caching Attribute Values Recipe 20.5. Using One Method as Accessorfor Multiple Attributes Recipe 20.6. Adding Functionality to a Class by Wrapping a Method Recipe 20.7. Adding Functionality to a Class by Enriching All Methods Recipe 20.8. Adding a Method to a Class Instance at Runtime Recipe 20.9. Checking Whether Interfaces Are Implemented Recipe 20.10. Using _ _new_ _ and _ _init_ _ Appropriately in Custom Metaclasses Recipe 20.11. Allowing Chaining of Mutating List Methods Recipe 20.12. Using Cooperative Super calls with Terser Syntax Recipe 20.13. Initializing Instance Attributes Without Using _ _init_ _ Recipe 20.14. Automatic Initialization of Instance Attributes Recipe 20.15. Upgrading Class Instances Automatically on reload Recipe 20.16. Binding Constants at Compile Time Recipe 20.17. Solving Metaclass Conflicts Colophon Index Copyright © 2005, 2002 O'Reilly Media, Inc. All rights reserved. Printed in the United States of America. Published by O'Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472. O'Reilly books may be purchased for educational, business, or sales promotional use. Online editions are also available for most titles (http://safari.oreilly.com). For more information, contact our corporate/institutional sales department: (800) 998-9938 or corporate@oreilly.com. Nutshell Handbook, the Nutshell Handbook logo, and the O'Reilly logo are registered trademarks of O'Reilly Media, Inc. The Cookbook series designations, Python Cookbook, the image of a springhaas, and related trade dress are trademarks of O'Reilly Media, Inc. Many of the designations used by manufacturers and sellers to distinguish their products are claimed as trademarks. Where those designations appear in this book, and O'Reilly Media, Inc. was aware of a trademark claim, the designations have been printed in caps or initial caps. While every precaution has been taken in the preparation of this book, the publisher and authors assume no responsibility for errors or omissions, or for damages resulting from the use of the information contained herein. Preface This book is not a typical O'Reilly book, written as a cohesive manuscript by one or two authors. Instead, it is a new kind of booka bold attempt at applying some principles of open source development to book authoring. Over 300 members of the Python community contributed materials to this book. In this Preface, we, the editors, want to give you, the reader, some background regarding how this book came about and the processes and people involved, and some thoughts about the implications of this new form. The Design of the Book In early 2000, Frank Willison, then Editor-in-Chief of O'Reilly & Associates, contacted me (David Ascher) to find out if I wanted to write a book. Frank had been the editor for Learning Python, which I cowrote with Mark Lutz. Since I had just taken a job at what was then considered a Perl shop (ActiveState), I didn't have the bandwidth necessary to write another book, and plans for the project were gently shelved. Periodically, however, Frank would send me an email or chat with me at a conference regarding some of the book topics we had discussed. One of Frank's ideas was to create a Python Cookbook, based on the concept first used by Tom Christiansen and Nathan Torkington with the Perl Cookbook. Frank wanted to replicate the success of the Perl Cookbook, but he wanted a broader set of people to provide input. He thought that, much as in a real cookbook, a larger set of authors would provide for a greater range of tastes. The quality, in his vision, would be ensured by the oversight of a technical editor, combined with O'Reilly's editorial review process. Frank and Dick Hardt, ActiveState's CEO, realized that Frank's goal could be combined with ActiveState's goal of creating a community site for open source programmers, called the ActiveState Programmer's Network (ASPN). ActiveState had a popular web site, with the infrastructure required to host a wide variety of content, but it wasn't in the business of creating original content. ActiveState always felt that the open source communities were the best sources of accurate and up-to-date content, even if sometimes that content was hard to find. The O'Reilly and ActiveState teams quickly realized that the two goals were aligned and that a joint venture would be the best way to achieve the following key objectives: Creating an online repository of Python recipes by Python programmers for Python programmers Publishing a book containing the best of those recipes, accompanied by overviews and background material written by key Python figures Learning what it would take to create a book with a different authoring model At the same time, two other activities were happening. First, those of us at ActiveState, including Paul Prescod, were actively looking for "stars" to join ActiveState's development team. One of the candidates being recruited was the famous (but unknown to us, at the time) Alex Martelli. Alex was famous because of his numerous and exhaustive postings on the Python mailing list, where he exhibited an unending patience for explaining Python's subtleties and joys to the increasing audience of Python programmers. He was unknown because he lived in Italy and, since he was a relative newcomer to the Python community, none of the old Python hands had ever met himtheir paths had not happened to cross back in the 1980s when Alex lived in the United States, working for IBM Research and enthusiastically using and promoting other high- level languages (at the time, mostly IBM's Rexx). ActiveState wooed Alex, trying to convince him to move to Vancouver. We came quite close, but his employer put some golden handcuffs on him, and somehow Vancouver's weather couldn't compete with Italy's. Alex stayed in Italy, much to my disappointment. As it happened, Alex was also at that time negotiating with O'Reilly about writing a book. Alex wanted to write a cookbook, but O'Reilly explained that the cookbook was already signed. Later, Alex and O'Reilly signed a contract for Python in Nutshell. The second ongoing activity was the creation of the Python Software Foundation. For a variety of reasons, best left to discussion over beers at a conference, everyone in the Python community wanted to create a non-profit organization that would be the holder of Python's intellectual property, to ensure that Python would be on a legally strong footing. However, such an organization needed both financial support and buy-in from the Python community to besuccessful. Given all these parameters, the various parties agreed to the following plan: ActiveState would build an online cookbook, a mechanism by which anyone could submit a recipe (i.e., a snippet of Python code addressing a particular problem, accompanied by a discussion of the recipe, much like a description of why one should use cream of tartar when whipping egg whites). To foster a community of authors and encourage peer review, the web site would also let readers of the recipes suggest changes, ask questions, and so on. As part of my ActiveState job, I would edit and ensure the quality of the recipes. Alex Martelli joined the project as a co-editor when the material was being prepared for publication, and, with Anna Martelli Ravenscroft, took over as primary editor for the second edition. O'Reilly would publish the best recipes as the Python Cookbook. In lieu of author royalties for the recipes, a portion of the proceeds from the book sales would be donated to the Python Software Foundation. The Implementation of the Book The online cookbook (at http://aspn.activestate.com/ASPN/Cookbook/Python/) was the entry point for the recipes. Users got free accounts, filled in a form, and presto, their recipes became part of the cookbook. Thousands of people read the recipes, and some added comments, and so, in the publishing equivalent of peer review, the recipes matured and grew. While it was predictable that the chance of getting your name in print would get people attracted to the online cookbook, the ongoing success of the cookbook, with dozens of recipes added monthly and more and more references to it on the newsgroups, is a testament to the value it brings to the readersvalue which is provided by the recipe authors. Starting from the materials available on the site, the implementation of the book was mostly a question of selecting, merging, ordering, and editing the materials. A few more details about this part of the work are in the "Organization" section of this Preface. Using the Code from This Book This book is here to help you get your job done. In general, you may use the code in this book in your programs and documentation. You do not need to contact us for permission unless you're reproducing a significant portion of the code. For example, writing a program that uses several chunks of code from this book does not require permission. Selling or distributing a CD-ROM of code taken from O'Reilly books does require permission. Answering a question by citing this book and quoting example code does not require permission. Incorporating a significant amount of code from this book into your product's documentation does require permission. We appreciate, but do not require, attribution. An attribution usually includes the title, author, publisher, and ISBN. For example: "Python Cookbook, 2d ed., by Alex Martelli, Anna Martelli Ravenscroft, and David Ascher (O'Reilly Media, 2005) 0-596-00797-3." If you feel your use of code from this book falls outside fair use or the permission given above, feel free to contact us at permissions@oreilly.com. Audience We expect that you know at least some Python. This book does not attempt to teach Python as a whole; rather, it presents some specific techniques and concepts (and occasionally tricks) for dealing with particular tasks. If you are looking for an introduction to Python, consider some of the books described in the Further Reading section of this Preface. However, you don't need to know a lot of Python to find this book helpful. Chapters include recipes demonstrating the best techniques for accomplishing some elementary and general tasks, as well as more complex or specialized ones. We have also added sidebars, here and there, to clarify certain concepts which are used in the book and which you may have heard of, but which might still be unclear to you. However, this is definitely not a book just for beginners. The main target audience is the whole Python community, mostly made up of pretty good programmers, neither newbies nor wizards. And if you do already know a lot about Python, you may be in for a pleasant surprise! We've included recipes that explore some the newest and least well-known areas of Python. You might very well learn a few thingswe did! Regardless of where you fall along the spectrum of Python expertise, and more generally of programming skill, we believe you will get something valuable from this book. If you already own the first edition, you may be wondering whether you need this second edition, too. We think the answer is "yes." The first edition had 245 recipes; we kept 146 of those (with lots of editing in almost all cases), and added 192 new ones, for a total of 338 recipes in this second edition. So, over half of the recipes in this edition are completely new, and all the recipes are updated to apply to today's Pythonreleases 2.3 and 2.4. Indeed, this update is the main factor which lets us have almost 100 more recipes in a book of about the same size. The first edition covered all versions from 1.5.2 (and sometimes earlier) to 2.2; this one focuses firmly on 2.3 and 2.4. Thanks to the greater power of today's Python, and, even more, thanks to the fact that this edition avoids the "historical" treatises about how you had to do things in Python versions released 5 or more years ago, we were able to provide substantially more currently relevant recipes and information in roughly the same amount of space. Organization This book has 20 chapters. Each chapter is devoted to a particular kind of recipe, such as algorithms, text processing, databases, and so on. The 1st edition had 17 chapters. There have been improvements to Python, both language and library, and to the corpus of recipes the Python community has posted to the cookbook site, that convinced us to add three entirely new chapters: on the iterators and generators introduced in Python 2.3; on Python's support for time and money operations, both old and new; and on new, advanced tools introduced in Python 2.2 and following releases (custom descriptors, decorators, metaclasses). Each chapter contains an introduction, written by an expert in the field, followed by recipes selected from the online cookbook (in some casesabout 5% of this book's recipesa few new recipes were specially written for this volume) and edited to fit the book's formatting and style requirements. Alex (with some help from Anna) did the vast majority of the selectiondetermining which recipes from the first edition to keep and update, and selecting new recipes to add, or merge with others, from the nearly 1,000 available on the site (so, if a recipe you posted to the cookbook site didn't get into this printed edition, it's his fault!). He also decided which subjects just had to be covered and thus might need specially written recipesalthough he couldn't manage to get quite all of the specially written recipes he wanted, so anything that's missing, and wasn't on the cookbook site, might not be entirely his fault. Once the selection was complete, the work turned to editing the recipes, and to merging multiple recipes, as well as incorporating important contents from many significant comments posted about the recipes. This proved to be quite a challenge, just as it had been for the first edition, but even more so. The recipes varied widely in their organization, level of completeness, and sophistication. With over 300 authors involved, over 300 different "voices" were included in the text. We have striven to maintain a variety of styles to reflect the true nature of this book, the book written by the entire Python community. However, we edited each recipe, sometimes quite considerably, to make it as accessible and useful as possible, ensuring enough uniformity in structure and presentation to maximize the usability of the book as a whole. Most recipes, both from the first edition and from the online site, had to be updated, sometimes heavily, to take advantage of new tools and better approaches developed since those recipes were originally posted. We also carefully reconsidered (and slightly altered) the ordering of chapters, and the placement and ordering of recipes within chapters; our goal in this reordering was to maximize the book's usefulness for both newcomers to Python and seasoned veterans, and, also, for both readers tackling the book sequentially, cover to cover, and ones just dipping in, in "random access" fashion, to look for help on some specific area. While the book should thus definitely be accessible "by hops and jumps," we nevertheless believe a first sequential skim will amply repay the modest time you, the reader, invest in it. On such a skim, skip every recipe that you have trouble following or that is of no current interest to you. Despite the skipping, you'll still get a sense of how the whole book hangs together and of where certain subjects are covered, which will stand you in good stead both for later in-depth sequential reading, if that's your choice, and for "random access" reading. To further help you get a sense of what's where in the book, here's a capsule summary of each chapter's contents, and equally capsule bios of the Python experts who were so kind as to take on the task of writing the chapters' "Introduction" sections. Chapter 1, introduction by Fred L. Drake, Jr. This chapter contains recipes for manipulating text in a variety of ways, including combining, filtering, and formatting strings, substituting variables throughout a text document, and dealing with Unicode.Fred Drake is a member of the PythonLabs group, working on Python development. A father of three, Fred is best known in the Python community for single-handedly maintaining the official documentation. Fred is a co-author of Python & XML (O'Reilly). Chapter 2, introduction by Mark Lutz This chapter presents techniques for working with data in files and for manipulating files and directories within the filesystem, including specific file formats and archive formats such as tar and zip. Mark Lutz is well known to most Python users as the most prolific author of Python books, including Programming Python, Python Pocket Reference, and Learning Python (all from O'Reilly), which he co-authored with David Ascher. Mark is also a leading Python trainer, spreading the Python gospel throughout the world. Chapter 3, introduction by Gustavo Niemeyer and Facundo Batista This chapter (new in this edition) presents tools and techniques for working with dates, times, decimal numbers, and some other money-related issues. Gustavo Niemeyer is the author of the third-party dateutil module, as well as a variety of other Python extensions and projects. Gustavo lives in Brazil. Facundo Batista is the author of the Decimal PEP 327, and of the standard library module decimal, which brought floating-point decimal support to Python 2.4. He lives in Argentina. The editors were delighted to bring them together for this introduction. Chapter 4, introduction by David Ascher This chapter includes recipes for many common techniques that can be used anywhere, or that don't really fit into any of the other, more specific recipe categories. David Ascher is a co-editor of this volume. David's background spans physics, vision research, scientific visualization, computer graphics, a variety of programming languages, co-authoring Learning Python (O'Reilly), teaching Python, and these days, a slew of technical and nontechnical tasks such as managing the ActiveState team. David also gets roped into organizing Python conferences on a regular basis. Chapter 5, introduction by Tim Peters This chapter covers techniques for searching and sorting in Python. Many of the recipes explore creative uses of the stable and fast list.sort in conjunction with the decorate- sort-undecorate (DSU) idiom (newly built in with Python 2.4), while others demonstrate the power of heapq, bisect, and other Python searching and sorting tools. Tim Peters, also known as the tim-bot, is one of the mythological figures of the Python world. He is the oracle, channeling Guido van Rossum when Guido is busy, channeling the IEEE-754 floating-point committee when anyone asks anything remotely relevant, and appearing conservative while pushing for a constant evolution in the language. Tim is a member of the PythonLabs team. Chapter 6, introduction by Alex Martelli This chapter offers a wide range of recipes that demonstrate the power of object-orientedprogramming with Python, including fundamental techniques such as delegating and controlling attribute access via special methods, intermediate ones such as the implementation of various design patterns, and some simple but useful applications of advanced concepts, such as custom metaclasses, which are covered in greater depth in Chapter 20. Alex Martelli, also known as the martelli-bot, is a co-editor of this volume. After almost a decade with IBM Research, then a bit more than that with think3, inc., Alex now works as a freelance consultant, most recently for AB Strakt, a Swedish Python-centered firm. He also edits and writes Python articles and books, including Python in a Nutshell (O'Reilly) and, occasionally, research works on the game of contract bridge. Chapter 7, introduction by Aaron Watters This chapter presents Python techniques for persistence, including serialization approaches and interaction with various databases. Aaron Watters was one of the earliest advocates of Python and is an expert in databases. He's known for having been the lead author on the first book on Python (Internet Programming with Python, M&T Books, now out of print), and he has authored many widely used Python extensions, such as kjBuckets and kwParsing. Aaron currently works as a freelance consultant. Chapter 8, introduction by Mark Hammond This chapter includes a collection of recipes that assist with the debugging and testing process, from customizing error logging and traceback information, to unit testing with custom modules, unittest and doctest. Mark Hammond is best known for his work supporting Python on the Windows platform. With Greg Stein, he built an incredible library of modules interfacing Python to a wide variety of APIs, libraries, and component models such as COM. He is also an expert designer and builder of developer tools, most notably Pythonwin and Komodo. Finally, Mark is an expert at debugging even the most messy systemsduring Komodo development, for example, Mark was often called upon to debug problems that spanned three languages (Python, C++, JavaScript), multiple threads, and multiple processes. Mark is also co-author, with Andy Robinson, of Python Programming on Win32 (O'Reilly). Chapter 9, introduction by Greg Wilson This chapter covers a variety of techniques for concurrent programming, including threads, queues, and multiple processes. Greg Wilson writes children's books, as well as books on parallel programming and data crunching. When he's not doing that, he's a contributing editor with Doctor Dobb's Journal, an adjunct professor in Computer Science at the University of Toronto, and a freelance software developer. Greg was the original driving force behind the Software Carpentry project, and he recently received a grant from the Python Software Foundation to develop Pythonic course material for computational scientists and engineers. Chapter 10, introduction by Donn Cave This chapter includes recipes for a number of common system administration tasks, from generating passwords and interacting with the Windows registry, to handling mailbox andweb server issues. Donn Cave is a software engineer at the University of Washington's central computer site. Over the years, Donn has proven to be a fount of information on comp.lang.python on all matters related to system calls, Unix, system administration, files, signals, and the like. Chapter 11, introduction by Fredrik Lundh This chapter contains recipes for common GUI tasks, mostly with Tkinter, but also a smattering of wxPython, Qt, image processing, and GUI recipes specific to Jython (for JVMJava Virtual Machine), Mac OS X, and IronPython (for dotNET). Fredrik Lundh, also known as the eff-bot, is the CTO of Secret Labs AB, a Swedish Python- focused company providing a variety of products and technologies. Fredrik is the world's leading expert on Tkinter (the most popular GUI toolkit for Python), as well as the main author of the Python Imaging Library (PIL). He is also the author of Python Standard Library (O'Reilly), which is a good complement to this volume and focuses on the modules in the standard Python library. Finally, he is a prolific contributor to comp.lang.python, helping novices and experts alike. Chapter 12, introduction by Paul Prescod This chapter offers techniques for parsing, processing, and generating XML using a variety of Python tools. Paul Prescod is an expert in three technologies: Python, which he need not justify; XML, which makes sense in a pragmatic world (Paul is co-author of the XML Handbook, with Charles Goldfarb, published by Prentice Hall); and Unicode, which somehow must address some deep-seated desire for pain and confusion that neither of the other two technologies satisfies. Paul is currently a product manager at Blast Radius. Chapter 13, introduction by Guido van Rossum This chapter covers a variety of network programming techniques, from writing basic TCP clients and servers to manipulating MIME messages. Guido created Python, nurtured it throughout its infancy, and is shepherding its growth. Need we say more? Chapter 14, introduction by Andy McKay This chapter presents a variety of web-related recipes, including ones for CGI scripting, running a Java servlet with Jython, and accessing the content of web pages. Andy McKay is the co-founder and vice president of Enfold Systems. In the last few years, Andy went from being a happy Perl user to a fanatical Python, Zope, and Plone expert. He wrote the Definitive Guide to Plone (Apress) and runs the popular Zope discussion site, http://www.zopezen.org. Chapter 15, introduction by Jeremy Hylton This chapter provides recipes for using Python in simple distributed systems, including XML-RPC, CORBA, and Twisted's Perspective Broker.Jeremy Hylton works for Google. In addition to young twins, Jeremy's interests including programming language theory, parsers, and the like. As part of his work for CNRI, Jeremy worked on a variety of distributed systems. Chapter 16, introduction by Paul F. Dubois This chapter contains Python techniques that involve program introspection, currying, dynamic importing, distributing programs, lexing and parsing. Paul Dubois has been working at the Lawrence Livermore National Laboratory for many years, building software systems for scientists working on everything from nuclear simulations to climate modeling. He has considerable experience with a wide range of scientific computing problems, as well as experience with language design and advanced object-oriented programming techniques. Chapter 17, introduction by David Beazley This chapter offers techniques for extending Python and recipes that assist in the development of extensions. David Beazley's chief claim to fame is SWIG, an amazingly powerful hack that lets one quickly wrap C and other libraries and use them from Python, Tcl, Perl, and myriad other languages. Behind this seemingly language-neutral tool lies a Python supporter of the first order, as evidenced by his book, Python Essential Reference (New Riders). David Beazley is a fairly sick man (in a good way), leading us to believe that more scarily useful tools are likely to emerge from his brain. He's currently inflicting his sense of humor on computer science students at the University of Chicago. Chapter 18, introduction by Tim Peters This chapter provides a collection of fascinating and useful algorithms and data structures implemented in Python. See the discussion of Chapter 5 for information about Tim Peters. Chapter 19, introduction by Raymond Hettinger This chapter (new in this edition) contains recipes demonstrating the variety and power of iterators and generatorshow Python makes your loops' structures simpler, faster, and reusable. Raymond Hettinger is the creator of the itertools package, original proposer of generator expressions, and has become a major contributor to the development of Pythonif you don't know who originated and implemented some major novelty or important optimization in the 2.3 and 2.4 releases of Python, our advice is to bet it was Raymond! Chapter 20, introduction by Raymond Hettinger This chapter (new in this edition) provides an in-depth look into the infrastructural elements which make Python's OOP so powerful and smooth, and how you can exploit and customize them for fun and profit. From handy idioms for building properties, to aliasing and caching attributes, all the way to decorators which optimize your functions by hackingtheir bytecode and to a factory of custom metaclasses to solve metatype conflicts, this chapter shows how, while surely "there be dragons here," they're the wise, powerful and beneficent Chinese variety thereof...! See the discussion of Chapter 19 for information about Raymond Hettinger. Further Reading There are many texts available to help you learn Python or refine your Python knowledge, from introductory texts all the way to quite formal language descriptions. We recommend the following books for general information about Python (all these books cover at least Python 2.2, unless otherwise noted): Python Programming for the Absolute Beginner, by Michael Dawson (Thomson Course Technology), is a hands-on, highly accessible introduction to Python for people who have never programmed. Learning Python, by Mark Lutz and David Ascher (O'Reilly), is a thorough introduction to the fundamentals of Python. Practical Python, by Magnus Lie Hetland (APress), is an introduction to Python which also develops, in detail, ten fully worked out, substantial programs in many different areas. Dive into Python, by Mark Pilgrim (APress), is a fast-paced introduction to Python for experienced programmers, and it is also freely available for online reading and downloading (http://diveintopython.org/). Python Standard Library, by Fredrik Lundh (O'Reilly), provides a use case for each module in the rich library that comes with every standard Python distribution (in the current first edition, the book only covers Python up to 2.0). Programming Python, by Mark Lutz (O'Reilly), is a thorough rundown of Python programming techniques (in the current second edition, the book only covers Python up to 2.0). Python Essential Reference, by David Beazley (New Riders), is a quick reference that focuses on the Python language and the core Python libraries (in the current second edition, the book only covers Python up to 2.1). Python in a Nutshell, by Alex Martelli (O'Reilly), is a comprehensive quick reference to the Python language and the key libraries used by most Python programmers. In addition, several more special-purpose books can help you explore particular aspects of Python programming. Which books you will like best depends a lot on your areas of interest. From personal experience, the editors can recommend at least the following: Python and XML, by Christopher A. Jones and Fred L. Drake, Jr. (O'Reilly), offers thorough coverage of using Python to read, process, and transform XML. Jython Essentials, by Samuele Pedroni and Noel Rappin (O'Reilly), is the authoritative book on Jython, the port of Python to the JVM. Particularly useful if you already know some (or a lot of) Java. Game Programming with Python, by Sean Riley (Charles River Media), covers programming computer games with Python, all the way from advanced graphics to moderate amounts of "artificial intelligence." Python Web Programming, by Steve Holden (New Riders), covers building networked systems using Python, with introductions to many other related technologies (databases, HTTP, HTML, etc.). Very suitable for readers with none to medium experience with these fields, but has something to teach everyone.In addition to these books, other important sources of information can help explain some of the code in the recipes in this book. We've pointed out the information that seemed particularly relevant in the "See Also" sections of each recipe. In these sections, we often refer to the standard Python documentation: most often the Library Reference, sometimes the Reference Manual, and occasionally the Tutorial. This documentation is freely available in a variety of forms: On the python.org web site (at http://www.python.org/doc/), which always contains the most up-to-date documentation about Python. On the pydoc.org web site (at http://pydoc.org/), accompanied by module-by-module documentation of the standard library automatically generated by the very useful pydoc tool. In Python itself. Recent versions of Python boast a nice online help system, which is worth exploring if you've never used it. Just type help( ) at the interactive Python interpreter prompt to start exploring. As part of the online help in your Python installation. ActivePython's installer, for example, includes a searchable Windows help file. The standard Python distribution currently includes HTML pages, but there are plans to include a similar Windows Help file in future releases. We have not included specific section numbers in our references to the standard Python documentation, since the organization of these manuals can change from release to release. You should be able to use the table of contents and indexes to find the relevant material. For the Library Reference, in particular, the Module Index (an alphabetical list of all standard library modules, each module name being a hyperlink to the Library Reference documentation for that module) is invaluable. Similarly, we have not given specific pointers in our references to Python in a Nutshell: that book is still in its first edition (covering Python up to 2.2) at the time of this writing, but by the time you're reading, a second edition (covering Python 2.3 and 2.4) is likely to be forthcoming, if not already published. Conventions Used in This Book Pronouns: the first person singular is meant to convey that the recipe's or chapter introduction's author is speaking (when multiple credits are given for a recipe, the author is the first person credited); however, even such remarks have at times had to be edited enough that they may not reflect the original author's intended meaning (we, the editors, tried hard to avoid that, but we know we must have failed in some cases, since there were so many remarks, and authorial intent was often not entirely clear). The second person is meant to refer to you, the reader. The first person plural collectively indicates you, the reader, plus the recipe's author and co-authors, the editors, and my friend Joe (hi Joe!)in other words, it's a very inclusive "we" or "us." Code: each block of code may indicate a complete module or script (or, often, a Python source file that is usable both as a script and as a module), an isolated snippet from some hypothetical module or script, or part of a Python interactive interpreter session (indicated by the prompt >>>). The following typographical conventions are used throughout this book: Italic for commands, filenames, for emphasis, and for first use of a term. Constant width for general code fragments and keywords (mostly Python ones, but also other languages, such as C or HTML, where they occur). Constant width is also used for all names defined in Python's library and third-party modules. Constant width bold is used to emphasize particular lines within code listings and show output that is produced. How to Contact Us We have tested and verified all the information in this book to the best of our abilities, but you may find that some features have changed, or that we have let errors slip through the production of the book. Please let us know of any errors that you find, as well as any suggestions for future editions, by writing to: O'Reilly Media 1005 Gravenstein Highway North Sebastopol, CA 95472 (800) 998-9938 (in the United States or Canada) (707) 829-0515 (international/local) (707) 829-0104 (fax) We have a web site for the book, where we'll list examples, errata, and any plans for future editions. You can access this page at: http://www.oreilly.com/catalog/pythoncook2 To ask technical questions or comment on the book, send email to: bookquestions@oreilly.com For more information about our books, conferences, Resource Centers, and the O'Reilly Network, see our web site at: http://www.oreilly.com/ The online cookbook from which most of the recipes for this book were taken is available at: http://aspn.activestate.com/ASPN/Cookbook/Python Safari® Enabled When you see a Safari Enabled icon on the cover of your favorite technology book, that means the book is available online through the O'Reilly Network Safari Bookshelf. Safari offers a solution that's better than e-books. It's a virtual library that lets you easily search thousands of top tech books, cut and paste code samples, download chapters, and find quick answers when you need the most accurate, current information. Try it for free at http://safari.oreilly.com. Acknowledgments Most publications, from mysteries to scientific papers to computer books, claim that the work being published would not have been possible without the collaboration of many others, typically including local forensic scientists, colleagues, and children, respectively. This book makes this claim to an extreme degree. Most of the words, code, and ideas in this volume were contributed by people not listed on the front cover. The original recipe authors, readers who submitted useful and insightful comments to the cookbook web site, and the authors of the chapter introductions, are the true authors of the book, and they deserve the credit. David Ascher The software that runs the online cookbook was the product of Andy McKay's constant and diligent effort. Andy was ActiveState's key Zope developer during the online data-collection phase of this project, and one of the key developers behind ASPN (http://aspn.activestate.com), ActiveState's content site, which serves a wide variety of information for and by programmers of open source languages such as Python, Perl, PHP, Tcl, and XSLT. Andy McKay used to be a Perl developer, by the way. At about the same time that I started at ActiveState, the company decided to use Zope to build what would become ASPN. In the years that followed, Andy has become a Zope master and somewhat of a Python fanatic (without any advocacy from me!), and is currently a Zope and Plone author, consultant and entrepreneur. Based on an original design that I put together with Diane Mueller, also of ActiveState, Andy single-handedly implemented ASPN in record time, then proceeded to adjust it to ever-changing requirements for new features that we hadn't anticipated in the early design phase, staying cheerful and professional throughout. It's a pleasure to have him as the author of the introduction to the chapter on web recipes. Since Andy's departure, James McGill has taken over as caretaker of the online cookbookhe makes sure that the cookbook is live at all hours of the day or night, ready to serve Pythonistas worldwide. Paul Prescod, then also of ActiveState, was a kindred spirit throughout the project, helping with the online editorial process, suggesting changes, and encouraging readers of comp.lang.python to visit the web site and submit recipes. Paul also helped with some of his considerable XML knowledge when it came to figuring out how to take the data out of Zope and get it ready for the publication process. The last activator I'd like to thank, for two different reasons, is Dick Hardt, founder and CEO of ActiveState. The first is that Dick agreed to let me work on the cookbook as part of my job. Had he not, I wouldn't have been able to participate in it. The second reason I'd like to thank Dick is for suggesting at the outset that a share of the book royalties go to the Python Software Foundation. This decision not only made it easier to enlist Python users into becoming contributors but has also resulted in some long-term revenue to an organization that I believe needs and deserves financial support. All Python users will benefit. Writing a software system a second time is dangerous; the "second-system" syndrome is a well- known engineering scenario in which teams that are allowed to rebuild systems "right" often end up with interminable, over-engineered projects. I'm pleased to say that this didn't happen in the case of this second edition, for two primary reasons. The first was the decision to trim the scope of the cookbook to cover only truly modern Pythonthat made the content more manageable and the book much more interesting to contemporary audiences. The second factor was that everyone realized with hindsight that I would have no time to contribute to the day-to-day editing of this second edition. I'm as glad as ever to have been associated with this book, and pleased that I have no guilt regarding the amount of work I didn't contribute. When people like Alex and Anna are willing to take on the work, it's much better for everyone else to get out of the way.Finally, I'd like to thank the O'Reilly editors who have had a big hand in shaping the cookbook. Laura Lewin was the original editor for the first edition, and she helped make sure that the project moved along, securing and coordinating the contributions of the introduction authors. Paula Ferguson then took the baton, provided a huge amount of precious feedback, and copyedited the final manuscript, ensuring that the prose was as readable as possible given the multiplicity of voices in the book. Jonathan Gennick was the editor for the second edition, and as far as I can tell, he basically let Alex and Anna drive, which was the right thing to do. Another editor I forgot to mention last time was Tim O'Reilly, who got more involved in this book than in most, in its early (rough) phases, and provided very useful input. Each time I review this acknowledgments section, I can't help but remember O'Reilly's Editor-in- Chief at the inception of the project, Frank Willison. Frank died suddenly on a black day, July 30, 2001. He was the person who most wanted to see this book happen, for the simple reason that he believed the Python community deserved it. Frank was always willing to explore new ideas, and he was generous to a fault. The idea of a book with over a hundred authors would have terrified most editors. Frank saw it as a challenge and an experiment. I still miss Frank. Alex Martelli I first met Python thanks to the gentle insistence of a former colleague, Alessandro Bottoni. He kept courteously repeating that I really should give Python a try, in spite of my claims that I already knew more programming languages than I knew what to do with. If I hadn't trusted his technical and aesthetic judgment enough to invest the needed time and energy on the basis of his suggestion, I most definitely wouldn't be writing and editing Python books today. Thanks for your well-placed stubbornness, Alessandro! Of course, once I tasted Python, I was irretrievably hookedmy lifelong taste for very high-level (often mis-named "scripting") languages at last congealed into one superb synthesis. Here, at long last, was a language with the syntactic ease of Rexx (and then some), the semantic simplicity of Tcl (and then some), the intellectual rigor of Scheme (and other Lisp variants), and the awesome power of Perl (and then some). How could I resist? Still, I do owe a debt to Mike Cowlishaw (inventor of Rexx), who I had the pleasure of having as a colleague when I worked for IBM Research, for first getting me hooked on scripting. I must also thank John Ousterhout and Larry Wall, the inventors of Tcl and Perl, respectively, for later reinforcing my addiction through their brainchildren. Greg Wilson first introduced me to O'Reilly, so he must get his share of thanks, tooand I'm overjoyed at having him as one of the introduction authors. I am also grateful to David Ascher, and several people at O'Reilly, for signing me up as co-editor of the first edition of this book and supporting so immediately and enthusiastically my idea that, hmmm, the time had sure come for a second edition (in dazed retrospect, I suspect what I meant was mostly that I had forgotten how deuced much work it had been to do the first one . . . and failed to realize that, with all the new materials heaped on ActiveState's site, as well as Python's wonderful progress over three years, the second edition would take more work than the first one. . . !). I couldn't possibly have done the job without an impressive array of technology to help me. I don't know the names of all the people I should thank for the Internet, ADSL, and Google's search engines, which, together, let me look things up so easilyor for many of the other hardware and software technologies cooperating to amplify my productivity. But, I do know I couldn't have made it without Theo de Raadt's OpenBSD operating system, Steve Jobs' inspiration behind Mac OS X and the iBook G4 on which I did most of the work, Bram Moolenaar's VIM editor, and, of course, Guido van Rossum's Python language. So, I'll single out Theo, Steve, Bram, and Guido for special thanks! Nor, as any book author will surely confirm, could I have done it without patience and moral support from friends and familychiefly my children Lucio and Flavia, my sister Elisabetta, my father Lanfranco. But the one person who was truly indispensable to this second edition was my wife and co-editor Anna. Having reconnected (after many years apart) thanks to Python, taken our honeymoon at the Open Source Convention, given a joint Lightning Talk about our "PythonicMarriage," maybe I should have surmised how wonderful it would be to work so closely with her, day in and day out, on such a large and complex joint project. It was truly incredible, all the way through, fully including the heated debates about this or that technical or organizational point or exact choice of wording in delicate cases. Throughout the effort and the stress, her skill, her love, her joy, always shined through, sustained me, and constantly renewed my energies and my determination. Thanks, Anna! Anna Martelli Ravenscroft I discovered Python about two years ago. I fell in love, both with Python and (concurrently) with the martelli-bot. Python is a language that is near to my heart, primarily because it is so quickly usable. It doesn't require you to become a hermit for the next four years in order to do anything with the language. Thank you to Guido. And thanks to the amazing Python community for providing such a welcoming atmosphere to newcomers. Working on this book was quite the learning experience for me. Besides all the Python code, I also learned both XML and VI, as well as reacquainting myself with Subversion. Thanks go to Holger Krekel and codespeak, for hosting our subversion repository while we travelled. Which brings us to a group of people who deserve special thanks: our reviewers. Holger Krekel, again, was exceptionally thorough, and ensured, among other things, that we had solid Unicode support. Raymond Hettinger gave us a huge amount of valuable, detailed insight throughout, particularly where iterators and generators were concerned. Both Raymond and Holger often offered alternatives to the presented "solutions" when warranted. Valentino Volonghi pointed out programming style issues as well as formatting issues and brought an incredible amount of enthusiasm to his reviews. Ryan Alexander, a newcomer to Python with a background in Java, provided extremely detailed recommendations on ordering and presenting materials (recipes and chapters), as well as pointing out explanations that were weak or missing altogether. His perspective was invaluable in making this book more accessible and useful to new Pythonistas. Several other individuals provided feedback on specific chapters or recipes, too numerous to list here. Your work, however, is greatly appreciated. Of course, thanks go to my husband. I am amazed at Alex's patience with questions (and I questioned a lot). His dedication to excellence is a co-author's dream. When presented with feedback, he consistently responded with appreciation and focus on making the book better. He's one of the least ego-istical writers I've ever met. Thank you to Dan, for encouraging my geekiness by starting me on Linux, teaching me proper terminology for the stuff I was doing, and for getting me hooked on the Internet. And finally, an extra special thanks to my children, Inanna and Graeme, for their hugs, understanding, and support when I was in geekmode, particularly during the final push to complete the book. You guys are the best kids a mother could wish for. Chapter 1. Text Introduction Recipe 1.1. Processing a String One Character at a Time Recipe 1.2. Converting Between Characters and Numeric Codes Recipe 1.3. Testing Whether an Object Is String-like Recipe 1.4. Aligning Strings Recipe 1.5. Trimming Space from the Ends of a String Recipe 1.6. Combining Strings Recipe 1.7. Reversing a String by Words or Characters Recipe 1.8. Checking Whether a String Contains a Set of Characters Recipe 1.9. Simplifying Usage of Strings' translate Method Recipe 1.10. Filtering a String for a Set of Characters Recipe 1.11. Checking Whether a String Is Text or Binary Recipe 1.12. Controlling Case Recipe 1.13. Accessing Substrings Recipe 1.14. Changing the Indentation of a Multiline String Recipe 1.15. Expanding and Compressing Tabs Recipe 1.16. Interpolating Variables in a String Recipe 1.17. Interpolating Variables in a Stringin Python 2.4 Recipe 1.18. Replacing Multiple Patterns in a Single Pass Recipe 1.19. Checking a String for Any of Multiple Endings Recipe 1.20. Handling International Text with Unicode Recipe 1.21. Converting Between Unicode and Plain Strings Recipe 1.22. Printing Unicode Charactersto Standard Output Recipe 1.23. Encoding Unicode Data for XML and HTML Recipe 1.24. Making Some Strings Case-Insensitive Recipe 1.25. Converting HTML Documents to Texton a Unix Terminal Introduction Credit: Fred L. Drake, Jr., PythonLabs Text-processing applications form a substantial part of the application space for any scripting language, if only because everyone can agree that text processing is useful. Everyone has bits of text that need to be reformatted or transformed in various ways. The catch, of course, is that every application is just a little bit different from every other application, so it can be difficult to find just the right reusable code to work with different file formats, no matter how similar they are. What Is Text? Sounds like an easy question, doesn't it? After all, we know it when we see it, don't we? Text is a sequence of characters, and it is distinguished from binary data by that very fact. Binary data, after all, is a sequence of bytes. Unfortunately, all data enters our applications as a sequence of bytes. There's no library function we can call that will tell us whether a particular sequence of bytes represents text, although we can create some useful heuristics that tell us whether data can safely (not necessarily correctly) be handled as text. Recipe 1.11 shows just such a heuristic. Python strings are immutable sequences of bytes or characters. Most of the ways we create and process strings treat them as sequences of characters, but many are just as applicable to sequences of bytes. Unicode strings are immutable sequences of Unicode characters: transformations of Unicode strings into and from plain strings use codecs (coder-decoders) objects that embody knowledge about the many standard ways in which sequences of characters can be represented by sequences of bytes (also known as encodings and character sets). Note that Unicode strings do not serve double duty as sequences of bytes. Recipe 1.20, Recipe 1.21, and Recipe 1.22 illustrate the fundamentals of Unicode in Python. Okay, let's assume that our application knows from the context that it's looking at text. That's usually the best approach because that's where external input comes into play. We're looking at a file either because it has a well-known name and defined format (common in the "Unix" world) or because it has a well-known filename extension that indicates the format of the contents (common on Windows). But now we have a problem: we had to use the word format to make the previous paragraph meaningful. Wasn't text supposed to be simple? Let's face it: there's no such thing as "pure" text, and if there were, we probably wouldn't care about it (with the possible exception of applications in the field of computational linguistics, where pure text may indeed sometimes be studied for its own sake). What we want to deal with in our applications is information contained in text. The text we care about may contain configuration data, commands to control or define processes, documents for human consumption, or even tabular data. Text that contains configuration data or a series of commands usually can be expected to conform to a fairly strict syntax that can be checked before relying on the information in the text. Informing the user of an error in the input text is typically sufficient to deal with things that aren't what we were expecting. Documents intended for humans tend to be simple, but they vary widely in detail. Since they are usually written in a natural language, their syntax and grammar can be difficult to check, at best. Different texts may use different character sets or encodings, and it can be difficult or even impossible to tell which character set or encoding was used to create a text if that information is not available in addition to the text itself. It is, however, necessary to support proper representation of natural-language documents. Natural-language text has structure as well, but the structures are often less explicit in the text and require at least some understanding of thelanguage in which the text was written. Characters make up words, which make up sentences, which make up paragraphs, and still larger structures may be present as well. Paragraphs alone can be particularly difficult to locate unless you know what typographical conventions were used for a document: is each line a paragraph, or can multiple lines make up a paragraph? If the latter, how do we tell which lines are grouped together to make a paragraph? Paragraphs may be separated by blank lines, indentation, or some other special mark. See Recipe 19.10 for an example of reading a text file as a sequence of paragraphs separated by blank lines. Tabular data has many issues that are similar to the problems associated with natural-language text, but it adds a second dimension to the input format: the text is no longer linearit is no longer a sequence of characters, but rather a matrix of characters from which individual blocks of text must be identified and organized. Basic Textual Operations As with any other data format, we need to do different things with text at different times. However, there are still three basic operations: Parsing the data into a structure internal to our application Transforming the input into something similar in some way, but with changes of some kind Generating completely new data Parsing can be performed in a variety of ways, and many formats can be suitably handled by ad hoc parsers that deal effectively with a very constrained format. Examples of this approach include parsers for RFC 2822-style email headers (see the rfc822 module in Python's standard library) and the configuration files handled by the ConfigParser module. The netrc module offers another example of a parser for an application-specific file format, this one based on the shlex module. shlex offers a fairly typical tokenizer for basic languages, useful in creating readable configuration files or allowing users to enter commands to an interactive prompt. These sorts of ad hoc parsers are abundant in Python's standard library, and recipes using them can be found in Chapter 2 and Chapter 13. More formal parsing tools are also available for Python; they depend on larger add-on packages and are surveyed in the introduction to Chapter 16. Transforming text from one format to another is more interesting when viewed as text processing, which is what we usually think of first when we talk about text. In this chapter, we'll take a look at some ways to approach transformations that can be applied for different purposes. Sometimes we'll work with text stored in external files, and other times we'll simply work with it as strings in memory. The generation of textual data from application-specific data structures is most easily performed using Python's print statement or the write method of a file or file-like object. This is often done using a method of the application object or a function, which takes the output file as a parameter. The function can then use statements such as these: print >>thefile, sometext thefile.write(sometext) which generate output to the appropriate file. However, this isn't generally thought of as text processing, as here there is no input text to be processed. Examples of using both print and write can of course be found throughout this book. Sources of Text Working with text stored as a string in memory can be easy when the text is not too large.Operations that search the text can operate over multiple lines very easily and quickly, and there's no need to worry about searching for something that might cross a buffer boundary. Being able to keep the text in memory as a simple string makes it very easy to take advantage of the built-in string operations available as methods of the string object. File-based transformations deserve special treatment, because there can be substantial overhead related to I/O performance and the amount of data that must actually be stored in memory. When working with data stored on disk, we often want to avoid loading entire files into memory, due to the size of the data: loading an 80 MB file into memory should not be done too casually! When our application needs only part of the data at a time, working on smaller segments of the data can yield substantial performance improvements, simply because we've allowed enough space for our program to run. If we are careful about buffer management, we can still maintain the performance advantage of using a small number of relatively large disk read and write operations by working on large chunks of data at a time. File-related recipes are found in Chapter 12 . Another interesting source for textual data comes to light when we consider the network. Text is often retrieved from the network using a socket. While we can always view a socket as a file (using the makefile method of the socket object), the data that is retrieved over a socket may come in chunks, or we may have to wait for more data to arrive. The textual data may not consist of all data until the end of the data stream, so a file object created with makefile may not be entirely appropriate to pass to text-processing code. When working with text from a network connection, we often need to read the data from the connection before passing it along for further processing. If the data is large, it can be handled by saving it to a file as it arrives and then using that file when performing text-processing operations. More elaborate solutions can be built when the text processing needs to be started before all the data is available. Examples of parsers that are useful in such situations may be found in the htmllib and HTMLParser modules in the standard library. String Basics The main tool Python gives us to process text is stringsimmutable sequences of characters. There are actually two kinds of strings: plain strings, which contain 8-bit (ASCII) characters; and Unicode strings, which contain Unicode characters. We won't deal much with Unicode strings here: their functionality is similar to that of plain strings, except each character takes up 2 (or 4) bytes, so that the number of different characters is in the tens of thousands (or even billions), as opposed to the 256 different characters that make up plain strings. Unicode strings are important if you must deal with text in many different alphabets, particularly Asian ideographs. Plain strings are sufficient to deal with English or any of a limited set of non-Asian languages. For example, all western European alphabets can be encoded in plain strings, typically using the international standard encoding known as ISO-8859-1 (or ISO-8859-15, if you need the Euro currency symbol as well). In Python, you express a literal string (curiously more often known as a string literal) as: 'this is a literal string' "this is another string" String values can be enclosed in either single or double quotes. The two different kinds of quotes work the same way, but having both allows you to include one kind of quotes inside of a string specified with the other kind of quotes, without needing to escape them with the backslash character: 'isn\'t that grand' "isn't that grand" To have a string literal span multiple lines, you can use a backslash as the last character on theline, which indicates that the next line is a continuation: big = "This is a long string\ that spans two lines." You must embed newlines in the string if you want the string to output on two lines: big = "This is a long string\n\ that prints on two lines." Another approach is to enclose the string in a pair of matching triple quotes (either single or double): bigger = """ This is an even bigger string that spans three lines. """ Using triple quotes, you don't need to use the continuation character, and line breaks in the string literal are preserved as newline characters in the resulting Python string object. You can also make a string literal "raw" string by preceding it with an r or R: big = r"This is a long string\ with a backslash and a newline in it" With a raw string, backslash escape sequences are left alone, rather than being interpreted. Finally, you can precede a string literal with a u or U to make it a Unicode string: hello = u'Hello\u0020World' Strings are immutable, which means that no matter what operation you do on a string, you will always produce a new string object, rather than mutating the existing string. A string is a sequence of characters, which means that you can access a single character by indexing: mystr = "my string" mystr[0] # 'm' mystr[-2] # 'n' You can also access a portion of the string with a slice: mystr[1:4] # 'y s' mystr[3:] # 'string' mystr[-3:] # 'ing' Slices can be extended, that is, include a third parameter that is known as the stride or step of the slice: mystr[:3:-1] # 'gnirt' mystr[1::2] # 'ysrn' You can loop on a string's characters:for c in mystr: This binds c to each of the characters in mystr in turn. You can form another sequence: list(mystr) # returns ['m','y',' ','s','t','r','i','n','g'] You can concatenate strings by addition: mystr+'oid' # 'my stringoid' You can also repeat strings by multiplication: 'xo'*3 # 'xoxoxo' In general, you can do anything to a string that you can do to any other sequence, as long as it doesn't require changing the sequence, since strings are immutable. String objects have many useful methods. For example, you can test a string's contents with s.isdigit( ), which returns true if s is not empty and all of the characters in s are digits (otherwise, it returns False). You can produce a new modified string with a method call such as s.toupper( ), which returns a new string that is like s, but with every letter changed into its uppercase equivalent. You can search for a string inside another with haystack.count('needle'), which returns the number of times the substring 'needle' appears in the string haystack. When you have a large string that spans multiple lines, you can split it into a list of single-line strings with splitlines: list_of_lines = one_large_string.splitlines( ) You can produce the single large string again with join: one_large_string = '\n'.join(list_of_lines) The recipes in this chapter show off many methods of the string object. You can find complete documentation in Python's Library Reference and Python in a Nutshell. Strings in Python can also be manipulated with regular expressions, via the re module. Regular expressions are a powerful (but complicated) set of tools that you may already be familiar with from another language (such as Perl), or from the use of tools such as the vi editor and text- mode commands such as grep. You'll find a number of uses of regular expressions in recipes in the second half of this chapter. For complete documentation, see the Library Reference and Python in a Nutshell. J.E.F. Friedl, Mastering Regular Expressions (O'Reilly) is also recommended if you need to master this subjectPython's regular expressions are basically the same as Perl's, which Friedl covers thoroughly. Python's standard module string offers much of the same functionality that is available from string methods, packaged up as functions instead of methods. The string module also offers a few additional functions, such as the useful string.maketrans function that is demonstrated in a few recipes in this chapter; several helpful string constants (string.digits, for example, is '0123456789') and, in Python 2.4, the new class Template, for simple yet flexible formatting of strings with embedded variables, which as you'll see features in one of this chapter's recipes. The string-formatting operator, %, provides a handy way to put strings together and to obtain precisely formatted strings from such objects as floating-point numbers. Again, you'll find recipes in this chapter that show how to use % for your purposes. Python also has lots of standard and extension modules that perform special processing on strings of many kinds. This chapter doesn't cover such specialized resources, but Chapter 12 is, for example, entirely devoted to theimportant specialized subject of processing XML. Recipe 1.1. Processing a String One Character at a Time Credit: Luther Blissett Problem You want to process a string one character at a time. Solution You can build a list whose items are the string's characters (meaning that the items are strings, each of length of onePython doesn't have a special type for "characters" as distinct from strings). Just call the built-in list, with the string as its argument: thelist = list(thestring) You may not even need to build the list, since you can loop directly on the string with a for statement: for c in thestring: do_something_with(c) or in the for clause of a list comprehension: results = [do_something_with(c) for c in thestring] or, with exactly the same effects as this list comprehension, you can call a function on each character with the map built-in function: results = map(do_something, thestring) Discussion In Python, characters are just strings of length one. You can loop over a string to access each of its characters, one by one. You can use map for much the same purpose, as long as what you need to do with each character is call a function on it. Finally, you can call the built-in type list to obtain a list of the length-one substrings of the string (i.e., the string's characters). If what you want is a set whose elements are the string's characters, you can call sets.Set with the string as the argument (in Python 2.4, you can also call the built-in set in just the same way): import sets magic_chars = sets.Set('abracadabra') poppins_chars = sets.Set('supercalifragilisticexpialidocious') print ''.join(magic_chars & poppins_chars) # set intersection acrdSee Also The Library Reference section on sequences; Perl Cookbook Recipe 1.5. Recipe 1.2. Converting Between Characters and Numeric Codes Credit: Luther Blissett Problem You need to turn a character into its numeric ASCII (ISO) or Unicode code, and vice versa. Solution That's what the built-in functions ord and chr are for: >>> print ord('a') 97 >>> print chr(97) a The built-in function ord also accepts as its argument a Unicode string of length one, in which case it returns a Unicode code value, up to 65536. To make a Unicode string of length one from a numeric Unicode code value, use the built-in function unichr: >>> print ord(u'\u2020') 8224 >>> print repr(unichr(8224)) u'\u2020' Discussion It's a mundane task, to be sure, but it is sometimes useful to turn a character (which in Python just means a string of length one) into its ASCII or Unicode code, and vice versa. The built-in functions ord, chr, and unichr cover all the related needs. Note, in particular, the huge difference between chr(n) and str(n), which beginners sometimes confuse...: >>> print repr(chr(97)) 'a' >>> print repr(str(97)) '97' chr takes as its argument a small integer and returns the corresponding single-character string according to ASCII, while str, called with any integer, returns the string that is the decimal representation of that integer. To turn a string into a list of character value codes, use the built-in functions map and ord together, as follows: >>> print map(ord, 'ciao') [99, 105, 97, 111]To build a string from a list of character codes, use ''.join, map and chr; for example: >>> print ''.join(map(chr, range(97, 100))) abc See Also Documentation for the built-in functions chr, ord, and unichr in the Library Reference and Python in a Nutshell. Recipe 1.3. Testing Whether an Object Is String-like Credit: Luther Blissett Problem You need to test if an object, typically an argument to a function or method you're writing, is a string (or more precisely, whether the object is string-like). Solution A simple and fast way to check whether something is a string or Unicode object is to use the built-ins isinstance and basestring, as follows: def isAString(anobj): return isinstance(anobj, basestring) Discussion The first approach to solving this recipe's problem that comes to many programmers' minds is type-testing: def isExactlyAString(anobj): return type(anobj) is type('') However, this approach is pretty bad, as it willfully destroys one of Python's greatest strengthssmooth, signature-based polymorphism. This kind of test would reject Unicode objects, instances of user-coded subclasses of str, and instances of any user-coded type that is meant to be "string-like". Using the isinstance built-in function, as recommended in this recipe's Solution, is much better. The built-in type basestring exists exactly to enable this approach. basestring is a common base class for the str and unicode types, and any string-like type that user code might define should also subclass basestring, just to make sure that such isinstance testing works as intended. basestring is essentially an "empty" type, just like object, so no cost is involved in subclassing it. Unfortunately, the canonical isinstance checking fails to accept such clearly string-like objects as instances of the UserString class from Python Standard Library module UserString, since that class, alas, does not inherit from basestring. If you need to support such types, you can check directly whether an object behaves like a stringfor example: def isStringLike(anobj): try: anobj + '' except: return False else: return True This isStringLike function is slower and more complicated than the isAString function presented in the "Solution", but it does accept instances of UserString (and other string-liketypes) as well as instances of str and unicode. The general Python approach to type-checking is known as duck typing: if it walks like a duck and quacks like a duck, it's duck-like enough for our purposes. The isStringLike function in this recipe goes only as far as the quacks-like part, but that may be enough. If and when you need to check for more string-like features of the object anobj, it's easy to test a few more properties by using a richer expression in the TRy clausefor example, changing the clause to: try: anobj.lower( ) + anobj + '' In my experience, however, the simple test shown in the isStringLike function usually does what I need. The most Pythonic approach to type validation (or any validation task, really) is just to try to perform whatever task you need to do, detecting and handling any errors or exceptions that might result if the situation is somehow invalidan approach known as "it's easier to ask forgiveness than permission" (EAFP). try/except is the key tool in enabling the EAFP style. Sometimes, as in this recipe, you may choose some simple task, such as concatenation to the empty string, as a stand-in for a much richer set of properties (such as, all the wealth of operations and methods that string objects make available). See Also Documentation for the built-ins isinstance and basestring in the Library Reference and Python in a Nutshell. Recipe 1.4. Aligning Strings Credit: Luther Blissett Problem You want to align strings: left, right, or center. Solution That's what the ljust, rjust, and center methods of string objects are for. Each takes a single argument, the width of the string you want as a result, and returns a copy of the starting string with spaces added on either or both sides: >>> print '|', 'hej'.ljust(20), '|', 'hej'.rjust(20), '|', 'hej'.center(20), '|' | hej | hej | hej | Discussion Centering, left-justifying, or right-justifying text comes up surprisingly oftenfor example, when you want to print a simple report with centered page numbers in a monospaced font. Because of this, Python string objects supply this functionality through three of their many methods. In Python 2.3, the padding character is always a space. In Python 2.4, however, while space- padding is still the default, you may optionally call any of these methods with a second argument, a single character to be used for the padding: >>> print 'hej'.center(20, '+') ++++++++hej+++++++++ See Also The Library Reference section on string methods; Java Cookbook recipe 3.5. Recipe 1.5. Trimming Space from the Ends of a String Credit: Luther Blissett Problem You need to work on a string without regard for any extra leading or trailing spaces a user may have typed. Solution That's what the lstrip, rstrip, and strip methods of string objects are for. Each takes no argument and returns a copy of the starting string, shorn of whitespace on either or both sides: >>> x = ' hej ' >>> print '|', x.lstrip( ), '|', x.rstrip( ), '|', x.strip( ), '|' | hej | hej | hej | Discussion Just as you may need to add space to either end of a string to align that string left, right, or center in a field of fixed width (as covered previously in Recipe 1.4), so may you need to remove all whitespace (blanks, tabs, newlines, etc.) from either or both ends. Because this need is frequent, Python string objects supply this functionality through three of their many methods. Optionally, you may call each of these methods with an argument, a string composed of all the characters you want to trim from either or both ends instead of trimming whitespace characters: >>> x = 'xyxxyy hejyx yyx' >>> print '|'+x.strip('xy')+'|' | hejyx | Note that in these cases the leading and trailing spaces have been left in the resulting string, as have the 'yx' that are followed by spaces: only all the occurrences of 'x' and 'y' at either end of the string have been removed from the resulting string. See Also The Library Reference section on string methods; Recipe 1.4; Java Cookbook recipe 3.12. Recipe 1.6. Combining Strings Credit: Luther Blissett Problem You have several small strings that you need to combine into one larger string. Solution To join a sequence of small strings into one large string, use the string operator join. Say that pieces is a list whose items are strings, and you want one big string with all the items concatenated in order; then, you should code: largeString = ''.join(pieces) To put together pieces stored in a few variables, the string-formatting operator % can often be even handier: largeString = '%s%s something %s yet more' % (small1, small2, small3) Discussion In Python, the + operator concatenates strings and therefore offers seemingly obvious solutions for putting small strings together into a larger one. For example, when you have pieces stored in a few variables, it seems quite natural to code something like: largeString = small1 + small2 + ' something ' + small3 + ' yet more' And similarly, when you have a sequence of small strings named pieces, it seems quite natural to code something like: largeString = '' for piece in pieces: largeString += piece Or, equivalently, but more fancifully and compactly: import operator largeString = reduce(operator.add, pieces, '') However, it's very important to realize that none of these seemingly obvious solution is goodthe approaches shown in the "Solution" are vastly superior. In Python, string objects are immutable. Therefore, any operation on a string, including string concatenation, produces a new string object, rather than modifying an existing one. Concatenating N strings thus involves building and then immediately throwing away each of N-1intermediate results. Performance is therefore vastly better for operations that build no intermediate results, but rather produce the desired end result at once. Python's string-formatting operator % is one such operation, particularly suitable when you have a few pieces (e.g., each bound to a different variable) that you want to put together, perhaps with some constant text in addition. Performance is not a major issue for this specific kind of task. However, the % operator also has other potential advantages, when compared to an expression that uses multiple + operations on strings. % is more readable, once you get used to it. Also, you don't have to call str on pieces that aren't already strings (e.g., numbers), because the format specifier %s does so implicitly. Another advantage is that you can use format specifiers other than %s, so that, for example, you can control how many significant digits the string form of a floating-point number should display. What Is "a Sequence?" Python does not have a specific type called sequence, but sequence is still an often- used term in Python. sequence, strictly speaking, means: a container that can be iterated on, to get a finite number of items, one at a time, and that also supports indexing, slicing, and being passed to the built-in function len (which gives the number of items in a container). Python lists are the "sequences" you'll meet most often, but there are many others (strings, unicode objects, tuples, array.arrays, etc.). Often, one does not need indexing, slicing, and lenthe ability to iterate, one item at a time, suffices. In that case, one should speak of an iterable (or, to focus on the finite number of items issue, a bounded iterable). Iterables that are not sequences include dictionaries (iteration gives the keys of the dictionary, one at a time in arbitrary order), file objects (iteration gives the lines of the text file, one at a time), and many more, including iterators and generators. Any iterable can be used in a for loop statement and in many equivalent contexts (the for clause of a list comprehension or Python 2.4 generator expression, and also many built-ins such as min, max, zip, sum, str.join, etc.). At http://www.python.org/moin/PythonGlossary, you can find a Python Glossary that can help you with these and several other terms. However, while the editors of this cookbook have tried to adhere to the word usage that the glossary describes, you will still find many places where this book says a sequence or an iterable or even a list, where, by strict terminology, one should always say a bounded iterable. For example, at the start of this recipe's Solution, we say "a sequence of small strings" where, in fact, any bounded iterable of strings suffices. The problem with using "bounded iterable" all over the place is that it would make this book read more like a mathematics textbook than a practical programming book! So, we have deviated from terminological rigor where readability, and maintaining in the book a variety of "voices", were better served by slightly imprecise terminology that is nevertheless entirely clear in context. When you have many small string pieces in a sequence, performance can become a truly important issue. The time needed to execute a loop using + or += (or a fancier but equivalent approach using the built-in function reduce) grows with the square of the number of characters you are accumulating, since the time to allocate and fill a large string is roughly proportional to the length of that string. Fortunately, Python offers an excellent alternative. The join method of a string object s takes as its only argument a sequence of strings and produces a string result obtained by concatenating all items in the sequence, with a copy of s joining each item to its neighbors. For example, ''.join(pieces) concatenates all the items of pieces in a single gulp,without interposing anything between them, and ', '.join(pieces) concatenates the items putting a comma and a space between each pair of them. It's the fastest, neatest, and most elegant and readable way to put a large string together. When the pieces are not all available at the same time, but rather come in sequentially from input or computation, use a list as an intermediate data structure to hold the pieces (to add items at the end of a list, you can call the append or extend methods of the list). At the end, when the list of pieces is complete, call ''.join(thelist) to obtain the big string that's the concatenation of all pieces. Of all the many handy tips and tricks I could give you about Python strings, I consider this one by far the most significant: the most frequent reason some Python programs are too slow is that they build up big strings with + or +=. So, train yourself never to do that. Use, instead, the ''.join approach recommented in this recipe. Python 2.4 makes a heroic attempt to ameliorate the issue, reducing a little the performance penalty due to such erroneous use of +=. While ''.join is still way faster and in all ways preferable, at least some newbie or careless programmer gets to waste somewhat fewer machine cycles. Similarly, psyco (a specializing just-in-time [JIT] Python compiler found at http://psyco.sourceforge.net/), can reduce the += penalty even further. Nevertheless, ''.join remains the best approach in all cases. See Also The Library Reference and Python in a Nutshell sections on string methods, string-formatting operations, and the operator module. Recipe 1.7. Reversing a String by Words or Characters Credit: Alex Martelli Problem You want to reverse the characters or words in a string. Solution Strings are immutable, so, to reverse one, we need to make a copy. The simplest approach for reversing is to take an extended slice with a "step" of -1, so that the slicing proceeds backwards: revchars = astring[::-1] To flip words, we need to make a list of words, reverse it, and join it back into a string with a space as the joiner: revwords = astring.split( ) # string -> list of words revwords.reverse( ) # reverse the list in place revwords = ' '.join(revwords) # list of strings -> string or, if you prefer terse and compact "one-liners": revwords = ' '.join(astring.split( )[::-1]) If you need to reverse by words while preserving untouched the intermediate whitespace, you can split by a regular expression: import re revwords = re.split(r'(\s+)', astring) # separators too, since '(...)' revwords.reverse( ) # reverse the list in place revwords = ''.join(revwords) # list of strings -> string Note that the joiner must be the empty string in this case, because the whitespace separators are kept in the revwords list (by using re.split with a regular expression that includes a parenthesized group). Again, you could make a one-liner, if you wished: revwords = ''.join(re.split(r'(\s+)', astring)[::-1]) but this is getting too dense and unreadable to be good Python code! Discussion In Python 2.4, you may make the by-word one-liners more readable by using the new built-in function reversed instead of the less readable extended-slicing indicator [::-1]:revwords = ' '.join(reversed(astring.split( ))) revwords = ''.join(reversed(re.split(r'(\s+)', astring))) For the by-character case, though, astring[::-1] remains best, even in 2.4, because to use reversed, you'd have to introduce a call to ''.join as well: revchars = ''.join(reversed(astring)) The new reversed built-in returns an iterator, suitable for looping on or for passing to some "accumulator" callable such as ''.joinit does not return a ready-made string! See Also Library Reference and Python in a Nutshell docs on sequence types and slicing, and (2.4 only) the reversed built-in; Perl Cookbook recipe 1.6. Recipe 1.8. Checking Whether a String Contains a Set of Characters Credit: Jürgen Hermann, Horst Hansen Problem You need to check for the occurrence of any of a set of characters in a string. Solution The simplest approach is clear, fast, and general (it works for any sequence, not just strings, and for any container on which you can test for membership, not just sets): def containsAny(seq, aset): """ Check whether sequence seq contains ANY of the items in aset. """ for c in seq: if c in aset: return True return False You can gain a little speed by moving to a higher-level, more sophisticated approach, based on the itertools standard library module, essentially expressing the same approach in a different way: import itertools def containsAny(seq, aset): for item in itertools.ifilter(aset._ _contains_ _, seq): return True return False Discussion Most problems related to sets are best handled by using the set built-in type introduced in Python 2.4 (if you're using Python 2.3, you can use the equivalent sets.Set type from the Python Standard Library). However, there are exceptions. Here, for example, a pure set-based approach would be something like: def containsAny(seq, aset): return bool(set(aset).intersection(seq)) However, with this approach, every item in seq inevitably has to be examined. The functions in this recipe's Solution, on the other hand, "short-circuit": they return as soon as they know the answer. They must still check every item in seq when the answer is Falsewe could never affirm that no item in seq is a member of aset without examining all the items, of course. But when the answer is true, we often learn about that very soon, namely as soon as we examine one item that is a member of aset. Whether this matters at all is very data-dependent, of course. It will make no practical difference when seq is short, or when the answer is typically False, but it may be extremely important for a very long seq (when the answer can typically be soon determinedto be true). The first version of containsAny presented in the recipe has the advantage of simplicity and clarity: it expresses the fundamental idea with total transparency. The second version may appear to be "clever", and that is not a complimentary adjective in the Python world, where simplicity and clarity are core values. However, the second version is well worth considering, because it shows a higher-level approach, based on the itertools module of the standard library. Higher-level approaches are most often preferable to lower-level ones (although the issue is moot in this particular case). itertools.ifilter takes a predicate and an iterable, and yields the items in that iterable that satisfy the "predicate". Here, as the "predicate", we use anyset._ _contains_ _, the bound method that is internally called when we code in anyset for membership testing. So, if ifilter yields anything at all, it yields an item of seq that is also a member of anyset, so we can return True as soon as this happens. If we get to the statement following the for, it must mean the return True never executed, because no items of seq are members of anyset, so we can return False. What Is "a Predicate?" A term you can see often in discussions about programming is predicate: it just means a function (or other callable object) that returns TRue or False as its result. A predicate is said to be satisfied when it returns true. If your application needs some function such as containsAny to check whether a string (or other sequence) contains any members of a set, you may also need such variants as: def containsOnly(seq, aset): """ Check whether sequence seq contains ONLY items in aset. """ for c in seq: if c not in aset: return False return True containsOnly is the same function as containsAny, but with the logic turned upside-down. Other apparently similar tasks don't lend themselves to short-circuiting (they intrinsically need to examine all items) and so are best tackled by using the built-in type set (in Python 2.4; in 2.3, you can use sets.Set in the same way): def containsAll(seq, aset): """ Check whether sequence seq contains ALL the items in aset. """ return not set(aset).difference(seq) If you're not accustomed to using the set (or sets.Set) method difference, be aware of its semantics: for any set a, a.difference(b) (just like a-set(b)) returns the set of all elements of a that are not in b. For example: >>> L1 = [1, 2, 3, 3] >>> L2 = [1, 2, 3, 4] >>> set(L1).difference(L2) set([ ]) >>> set(L2).difference(L1) set([4])which hopefully helps explain why: >>> containsAll(L1, L2) False >>> containsAll(L2, L1) True (In other words, don't confuse difference with another method of set, symmetric_difference, which returns the set of all items that are in either argument and not in the other.) When you're dealing specifically with (plain, not Unicode) strings for both seq and aset, you may not need the full generality of the functions presented in this recipe, and may want to try the more specialized approach explained in Recipe 1.10 based on strings' method TRanslate and the string.maketrans function from the Python Standard Library. For example: import string notrans = string.maketrans('', '') # identity "translation" def containsAny(astr, strset): return len(strset) != len(strset.translate(notrans, astr)) def containsAll(astr, strset): return not strset.translate(notrans, astr) This somewhat tricky approach relies on strset.translate(notrans, astr) being the subsequence of strset that is made of characters not in astr. When that subsequence has the same length as strset, no characters have been removed by strset.translate, therefore no characters of strset are in astr. Conversely, when the subsequence is empty, all characters have been removed, so all characters of strset are in astr. The translate method keeps coming up naturally when one wants to treat strings as sets of characters, because it's speedy as well as handy and flexible; see Recipe 1.10 for more details. These two sets of approaches to the recipe's tasks have very different levels of generality. The earlier approaches are very general: not at all limited to string processing, they make rather minimal demands on the objects you apply them to. The approach based on the translate method, on the other hand, works only when both astr and strset are strings, or very closely mimic plain strings' functionality. Not even Unicode strings suffice, because the TRanslate method of Unicode strings has a signature that is different from that of plain stringsa single argument (a dict mapping code numbers to Unicode strings or None) instead of two (both strings). See Also Recipe 1.10; documentation for the translate method of strings and Unicode objects, and maketrans function in the string module, in the Library Reference and Python in a Nutshell; ditto for documentation of built-in set (Python 2.4 only), modules sets and itertools, and the special method _ _contains_ _. Recipe 1.9. Simplifying Usage of Strings' translate Method Credit: Chris Perkins, Raymond Hettinger Problem You often want to use the fast code in strings' TRanslate method, but find it hard to remember in detail how that method and the function string.maketrans work, so you want a handy facade to simplify their use in typical cases. Solution The TRanslate method of strings is quite powerful and flexible, as detailed in Recipe 1.10. However, exactly because of that power and flexibility, it may be a nice idea to front it with a "facade" that simplifies its typical use. A little factory function, returning a closure, can do wonders for this kind of task: import string def translator(frm='', to='', delete='', keep=None): if len(to) == 1: to = to * len(frm) trans = string.maketrans(frm, to) if keep is not None: allchars = string.maketrans('', '') delete = allchars.translate(allchars, keep.translate(allchars, delete)) def translate(s): return s.translate(trans, delete) return translate Discussion I often find myself wanting to use strings' translate method for any one of a few purposes, but each time I have to stop and think about the details (see Recipe 1.10 for more information about those details). So, I wrote myself a class (later remade into the factory closure presented in this recipe's Solution) to encapsulate various possibilities behind a simpler-to-use facade. Now, when I want a function that keeps only characters from a given set, I can easily build and use that function: >>> digits_only = translator(keep=string.digits) >>> digits_only('Chris Perkins : 224-7992') '2247992' It's similarly simple when I want to remove a set of characters: >>> no_digits = translator(delete=string.digits) >>> no_digits('Chris Perkins : 224-7992') 'Chris Perkins : -'and when I want to replace a set of characters with a single character: >>> digits_to_hash = translator(from=string.digits, to='#') >>> digits_to_hash('Chris Perkins : 224-7992') 'Chris Perkins : ###-####' While the latter may appear to be a bit of a special case, it is a task that keeps coming up for me every once in a while. I had to make one arbitrary design decision in this recipenamely, I decided that the delete parameter "trumps" the keep parameter if they overlap: >>> trans = translator(delete='abcd', keep='cdef') >>> trans('abcdefg') 'ef' For your applications it might be preferable to ignore delete if keep is specified, or, perhaps better, to raise an exception if they are both specified, since it may not make much sense to let them both be given in the same call to translator, anyway. Also: as noted in Recipe 1.8 and Recipe 1.10, the code in this recipe works only for normal strings, not for Unicode strings. See Recipe 1.10 to learn how to code this kind of functionality for Unicode strings, whose translate method is different from that of plain (i.e., byte) strings. Closures A closure is nothing terribly complicated: just an "inner" function that refers to names (variables) that are local to an "outer" function containing it. Canonical toy- level example: def make_adder(addend): def adder(augend): return augend+addend return adder Executing p = make_adder(23) makes a closure of inner function adder internally referring to a name addend that is bound to the value 23. Then, q = make_adder(42) makes another closure, for which, internally, name addend is instead bound to the value 42. Making q in no way interferes with p, they can happily and independently coexist. So we can now execute, say, print p(100), q(100) and enjoy the output 123 142. In practice, you may often see make_adder referred to as a closure rather than by the pedantic, ponderous periphrasis "a function that returns a closure"fortunately, context often clarifies the situation. Calling make_adder a factory (or factory function) is both accurate and concise; you may also say it's a closure factory to specify it builds and returns closures, rather than, say, classes or class instances. See Also Recipe 1.10 for a direct equivalent of this recipe's TRanslator(keep=...), more information onthe TRanslate method, and an equivalent approach for Unicode strings; documentation for strings' translate method, and for the maketrans function in the string module, in the Library Reference and Python in a Nutshell. Recipe 1.10. Filtering a String for a Set of Characters Credit: Jürgen Hermann, Nick Perkins, Peter Cogolo Problem Given a set of characters to keep, you need to build a filtering function that, applied to any string s, returns a copy of s that contains only characters in the set. Solution The TRanslate method of string objects is fast and handy for all tasks of this ilk. However, to call translate effectively to solve this recipe's task, we must do some advance preparation. The first argument to TRanslate is a translation table: in this recipe, we do not want to do any translation, so we must prepare a first argument that specifies "no translation". The second argument to TRanslate specifies which characters we want to delete: since the task here says that we're given, instead, a set of characters to keep (i.e., to not delete), we must prepare a second argument that gives the set complementdeleting all characters we must not keep. A closure is the best way to do this advance preparation just once, obtaining a fast filtering function tailored to our exact needs: import string # Make a reusable string of all characters, which does double duty # as a translation table specifying "no translation whatsoever" allchars = string.maketrans('', '') def makefilter(keep): """ Return a function that takes a string and returns a partial copy of that string consisting of only the characters in 'keep'. Note that `keep' must be a plain string. """ # Make a string of all characters that are not in 'keep': the "set # complement" of keep, meaning the string of characters we must delete delchars = allchars.translate(allchars, keep) # Make and return the desired filtering function (as a closure) def thefilter(s): return s.translate(allchars, delchars) return thefilter if _ _name_ _ == '_ _main_ _': just_vowels = makefilter('aeiouy') print just_vowels('four score and seven years ago') # emits: ouoeaeeyeaao print just_vowels('tiger, tiger burning bright') # emits: ieieuii Discussion The key to understanding this recipe lies in the definitions of the maketrans function in the string module of the Python Standard Library and in the translate method of string objects. TRanslate returns a copy of the string you call it on, replacing each character in it with the corresponding character in the translation table passed in as the first argument and deleting thecharacters specified in the second argument. maketrans is a utility function to create translation tables. (A translation table is a string t of exactly 256 characters: when you pass t as the first argument of a translate method, each character c of the string on which you call the method is translated in the resulting string into the character t[ord(c)].) In this recipe, efficiency is maximized by splitting the filtering task into preparation and execution phases. The string of all characters is clearly reusable, so we build it once and for all as a global variable when this module is imported. That way, we ensure that each filtering function uses the same string-of-all-characters object, not wasting any memory. The string of characters to delete, which we need to pass as the second argument to the translate method, depends on the set of characters to keep, because it must be built as the "set complement" of the latter: we must tell translate to delete every character that we do not want to keep. So, we build the delete-these- characters string in the makefilter factory function. This building is done quite rapidly by using the translate method to delete the "characters to keep" from the string of all characters. The translate method is very fast, as are the construction and execution of these useful little resulting functions. The test code that executes when this recipe runs as a main script shows how to build a filtering function by calling makefilter, bind a name to the filtering function (by simply assigning the result of calling makefilter to a name), then call the filtering function on some strings and print the results. Incidentally, calling a filtering function with allchars as the argument puts the set of characters being kept into a canonic string form, alphabetically sorted and without duplicates. You can use this idea to code a very simple function to return the canonic form of any set of characters presented as an arbitrary string: def canonicform(s): """ Given a string s, return s's characters as a canonic-form string: alphabetized and without duplicates. """ return makefilter(s)(allchars) The Solution uses a def statement to make the nested function (closure) it returns, because def is the most normal, general, and clear way to make functions. If you prefer, you could use lambda instead, changing the def and return statements in function makefilter into just one return lambda statement: return lambda s: s.translate(allchars, delchars) Most Pythonistas, but not all, consider using def clearer and more readable than using lambda. Since this recipe deals with strings seen as sets of characters, you could alternatively use the sets.Set type (or, in Python 2.4, the new built-in set type) to perform the same tasks. Thanks to the translate method's power and speed, it's often faster to work directly on strings, rather than go through sets, for tasks of this ilk. However, just as noted in Recipe 1.8, the functions in this recipe only work for normal strings, not for Unicode strings. To solve this recipe's task for Unicode strings, we must do some very different preparation. A Unicode string's translate method takes only one argument: a mapping or sequence, which is indexed with the code number of each character in the string. Characters whose codes are not keys in the mapping (or indices in the sequence) are just copied over to the output string. Otherwise, the value corresponding to each character's code must be either a Unicode string (which is substituted for the character) or None (in which case the character is deleted). A very nice and powerful arrangement, but unfortunately not one that's identical to the way plain strings work, so we must recode. Normally, we use either a dict or a list as the argument to a Unicode string's translate method to translate some characters and/or delete some. But for the specific task of this recipe (i.e., keep just some characters, delete all others), we might need an inordinately large dict or string, just mapping all other characters to None. It's better to code, instead, a little class that appropriately implements a _ _getitem_ _ method (the special method that gets called inindexing operations). Once we're going to the (slight) trouble of coding a little class, we might as well make its instances callable and have makefilter be just a synonym for the class itself: import sets class Keeper(object): def _ _init_ _(self, keep): self.keep = sets.Set(map(ord, keep)) def _ _getitem_ _(self, n): if n not in self.keep: return None return unichr(n) def _ _call_ _(self, s): return unicode(s).translate(self) makefilter = Keeper if _ _name_ _ == '_ _main_ _': just_vowels = makefilter('aeiouy') print just_vowels(u'four score and seven years ago') # emits: ouoeaeeyeaao print just_vowels(u'tiger, tiger burning bright') # emits: ieieuii We might name the class itself makefilter, but, by convention, one normally names classes with an uppercase initial; there is essentially no cost in following that convention here, too, so we did. See Also Recipe 1.8; documentation for the TRanslate method of strings and Unicode objects, and maketrans function in the string module, in the Library Reference and Python in a Nutshell. Recipe 1.11. Checking Whether a String Is Text or Binary Credit: Andrew Dalke Problem Python can use a plain string to hold either text or arbitrary bytes, and you need to determine (heuristically, of course: there can be no precise algorithm for this) which of the two cases holds for a certain string. Solution We can use the same heuristic criteria as Perl does, deeming a string binary if it contains any nulls or if more than 30% of its characters have the high bit set (i.e., codes greater than 126) or are strange control codes. We have to code this ourselves, but this also means we easily get to tweak the heuristics for special application needs: from _ _future_ _ import division # ensure / does NOT truncate import string text_characters = "".join(map(chr, range(32, 127))) + "\n\r\t\b" _null_trans = string.maketrans("", "") def istext(s, text_characters=text_characters, threshold=0.30): # if s contains any null, it's not text: if "\0" in s: return False # an "empty" string is "text" (arbitrary but reasonable choice): if not s: return True # Get the substring of s made up of non-text characters t = s.translate(_null_trans, text_characters) # s is 'text' if less than 30% of its characters are non-text ones: return len(t)/len(s) <= threshold Discussion You can easily do minor customizations to the heuristics used by function istext by passing in specific values for the threshold, which defaults to 0.30 (30%), or for the string of those characters that are to be deemed "text" (which defaults to normal ASCII characters plus the four "normal" control characters, meaning ones that are often found in text). For example, if you expected Italian text encoded as ISO-8859-1, you could add the accented letters used in Italian, "àèéìÃ2Ã1", to the text_characters argument. Often, what you need to check as being either binary or text is not a string, but a file. Again, we can use the same heuristics as Perl, checking just the first block of the file with the istext function shown in this recipe's Solution: def istextfile(filename, blocksize=512, **kwds): return istext(open(filename).read(blocksize), **kwds)Note that, by default, the expression len(t)/len(s) used in the body of function istext would truncate the result to 0, since it is a division between integer numbers. In some future version (probably Python 3.0, a few years away), Python will change the meaning of the / operator so that it performs division without truncationif you really do want truncation, you should use the truncating-division operator, //. However, Python has not yet changed the semantics of division, keeping the old one by default in order to ensure backwards compatibility. It's important that the millions of lines of code of Python programs and modules that already exist keep running smoothly under all new 2.x versions of Pythononly upon a change of major language version number, no more often than every decade or so, is Python allowed to change in ways that aren't backwards-compatible. Since, in the small module containing this recipe's Solution, it's handy for us to get the division behavior that is scheduled for introduction in some future release, we start our module with the statement: from _ _future_ _ import division This statement doesn't affect the rest of the program, only the specific module that starts with this statement; throughout this module, / performs "true division" (without truncation). As of Python 2.3 and 2.4, division is the only thing you may want to import from _ _future_ _. Other features that used to be scheduled for the future, nested_scopes and generators, are now part of the language and cannot be turned offit's innocuous to import them, but it makes sense to do so only if your program also needs to run under some older version of Python. See Also Recipe 1.10 for more details about function maketrans and string method translate; Language Reference for details about true versus truncating division. Recipe 1.12. Controlling Case Credit: Luther Blissett Problem You need to convert a string from uppercase to lowercase, or vice versa. Solution That's what the upper and lower methods of string objects are for. Each takes no arguments and returns a copy of the string in which each letter has been changed to upper- or lowercase, respectively. big = little.upper( ) little = big.lower( ) Characters that are not letters are copied unchanged. s.capitalize is similar to s[:1].upper( )+s[1:].lower( ): the first character is changed to uppercase, and all others are changed to lowercase. s.title is again similar, but it capitalizes the first letter of each word (where a "word" is a sequence of letters) and uses lowercase for all other letters: >>> print 'one tWo thrEe'.capitalize( ) One two three >>> print 'one tWo thrEe'.title( ) One Two Three Discussion Case manipulation of strings is a very frequent need. Because of this, several string methods let you produce case-altered copies of strings. Moreover, you can also check whether a string object is already in a given case form, with the methods isupper, islower, and istitle, which all return true if the string is not empty, contains at least one letter, and already meets the uppercase, lowercase, or titlecase constraints. There is no analogous iscapitalized method, and coding it is not trivial, if we want behavior that's strictly similar to strings' is... methods. Those methods all return False for an "empty" string, and the three case-checking ones also return False for strings that, while not empty, contain no letters at all. The simplest and clearest way to code iscapitalized is clearly: def iscapitalized(s): return s == s.capitalize( ) However, this version deviates from the boundary-case semantics of the analogous is... methods, since it also returns TRue for strings that are empty or contain no letters. Here's a stricter one:import string notrans = string.maketrans('', '') # identity "translation" def containsAny(str, strset): return len(strset) != len(strset.translate(notrans, str)) def iscapitalized(s): return s == s.capitalize( ) and containsAny(s, string.letters) Here, we use the function shown in Recipe 1.8 to ensure we return False if s is empty or contains no letters. As noted in Recipe 1.8, this means that this specific version works only for plain strings, not for Unicode ones. See Also Library Reference and Python in a Nutshell docs on string methods; Perl Cookbook recipe 1.9; Recipe 1.8. Recipe 1.13. Accessing Substrings Credit: Alex Martelli Problem You want to access portions of a string. For example, you've read a fixed-width record and want to extract the record's fields. Solution Slicing is great, but it only does one field at a time: afield = theline[3:8] If you need to think in terms of field lengths, struct.unpack may be appropriate. For example: import struct # Get a 5-byte string, skip 3, get two 8-byte strings, then all the rest: baseformat = "5s 3x 8s 8s" # by how many bytes does theline exceed the length implied by this # base-format (24 bytes in this case, but struct.calcsize is general) numremain = len(theline) - struct.calcsize(baseformat) # complete the format with the appropriate 's' field, then unpack format = "%s %ds" % (baseformat, numremain) l, s1, s2, t = struct.unpack(format, theline) If you want to skip rather than get "all the rest", then just unpack the initial part of theline with the right length: l, s1, s2 = struct.unpack(baseformat, theline[:struct.calcsize(baseformat)]) If you need to split at five-byte boundaries, you can easily code a list comprehension (LC) of slices: fivers = [theline[k:k+5] for k in xrange(0, len(theline), 5)] Chopping a string into individual characters is of course easier: chars = list(theline) If you prefer to think of your data as being cut up at specific columns, slicing with LCs is generally handier: cuts = [8, 14, 20, 26, 30] pieces = [ theline[i:j] for i, j in zip([0]+cuts, cuts+[None]) ]The call to zip in this LC returns a list of pairs of the form (cuts[k], cuts[k+1]), except that the first pair is (0, cuts[0]), and the last one is (cuts[len(cuts)-1], None). In other words, each pair gives the right (i, j) for slicing between each cut and the next, except that the first one is for the slice before the first cut, and the last one is for the slice from the last cut to the end of the string. The rest of the LC just uses these pairs to cut up the appropriate slices of theline. Discussion This recipe was inspired by recipe 1.1 in the Perl Cookbook. Python's slicing takes the place of Perl's substr. Perl's built-in unpack and Python's struct.unpack are similar. Perl's is slightly richer, since it accepts a field length of * for the last field to mean all the rest. In Python, we have to compute and insert the exact length for either extraction or skipping. This isn't a major issue because such extraction tasks will usually be encapsulated into small functions. Memoizing, also known as automatic caching, may help with performance if the function is called repeatedly, since it allows you to avoid redoing the preparation of the format for the struct unpacking. See Recipe 18.5 for details about memoizing. In a purely Python context, the point of this recipe is to remind you that struct.unpack is often viable, and sometimes preferable, as an alternative to string slicing (not quite as often as unpack versus substr in Perl, given the lack of a *-valued field length, but often enough to be worth keeping in mind). Each of these snippets is, of course, best encapsulated in a function. Among other advantages, encapsulation ensures we don't have to work out the computation of the last field's length on each and every use. This function is the equivalent of the first snippet using struct.unpack in the "Solution": def fields(baseformat, theline, lastfield=False): # by how many bytes does theline exceed the length implied by # base-format (struct.calcsize computes exactly that length) numremain = len(theline)-struct.calcsize(baseformat) # complete the format with the appropriate 's' or 'x' field, then unpack format = "%s %d%s" % (baseformat, numremain, lastfield and "s" or "x") return struct.unpack(format, theline) A design decision worth noticing (and, perhaps, worth criticizing) is that of having a lastfield=False optional parameter. This reflects the observation that, while we often want to skip the last, unknown-length subfield, sometimes we want to retain it instead. The use of lastfield in the expression lastfield and s or x (equivalent to C's ternary operator lastfield? "s":"c") saves an if/else, but it's unclear whether the saving is worth the obscurity. See Recipe 18.9 for more about simulating ternary operators in Python. If function fields is called in a loop, memoizing (caching) with a key that is the tuple (baseformat, len(theline), lastfield) may offer faster performance. Here's a version of fields with memoizing: def fields(baseformat, theline, lastfield=False, _cache={ }): # build the key and try getting the cached format string key = baseformat, len(theline), lastfield format = _cache.get(key) if format is None: # no format string was cached, build and cache it numremain = len(theline)-struct.calcsize(baseformat) _cache[key] = format = "%s %d%s" % ( baseformat, numremain, lastfield and "s" or "x") return struct.unpack(format, theline)The idea behind this memoizing is to perform the somewhat costly preparation of format only once for each set of arguments requiring that preparation, thereafter storing it in the _cache dictionary. Of course, like all optimizations, memoizing needs to be validated by measuring performance to check that each given optimization does actually speed things up. In this case, I measure an increase in speed of approximately 30% to 40% for the memoized version, meaning that the optimization is probably not worth the bother unless the function is part of a performance bottleneck for your program. The function equivalent of the next LC snippet in the solution is: def split_by(theline, n, lastfield=False): # cut up all the needed pieces pieces = [theline[k:k+n] for k in xrange(0, len(theline), n)] # drop the last piece if too short and not required if not lastfield and len(pieces[-1]) < n: pieces.pop( ) return pieces And for the last snippet: def split_at(theline, cuts, lastfield=False): # cut up all the needed pieces pieces = [ theline[i:j] for i, j in zip([0]+cuts, cuts+[None]) ] # drop the last piece if not required if not lastfield: pieces.pop( ) return pieces In both of these cases, a list comprehension doing slicing turns out to be slightly preferable to the use of struct.unpack. A completely different approach is to use generators, such as: def split_at(the_line, cuts, lastfield=False): last = 0 for cut in cuts: yield the_line[last:cut] last = cut if lastfield: yield the_line[last:] def split_by(the_line, n, lastfield=False): return split_at(the_line, xrange(n, len(the_line), n), lastfield) Generator-based approaches are particularly appropriate when all you need to do on the sequence of resulting fields is loop over it, either explicitly, or implicitly by calling on it some "accumulator" callable such as ''.join. If you do need to materialize a list of the fields, and what you have available is a generator instead, you only need to call the built-in list on the generator, as in: list_of_fields = list(split_by(the_line, 5)) See Also Recipe 18.9 and Recipe 18.5; Perl Cookbook recipe 1.1. Recipe 1.14. Changing the Indentation of a Multiline String Credit: Tom Good Problem You have a string made up of multiple lines, and you need to build another string from it, adding or removing leading spaces on each line so that the indentation of each line is some absolute number of spaces. Solution The methods of string objects are quite handy, and let us write a simple function to perform this task: def reindent(s, numSpaces): leading_space = numSpaces * ' ' lines = [ leading_space + line.strip( ) for line in s.splitlines( ) ] return '\n'.join(lines) Discussion When working with text, it may be necessary to change the indentation level of a block. This recipe's code adds leading spaces to or removes them from each line of a multiline string so that the indentation level of each line matches some absolute number of spaces. For example: >>> x = """ line one ... line two ... and line three ... """ >>> print x line one line two and line three >>> print reindent(x, 4) line one line two and line three Even if the lines in s are initially indented differently, this recipe makes their indentation homogeneous, which is sometimes what we want, and sometimes not. A frequent need is to adjust the amount of leading spaces in each line, so that the relative indentation of each line in the block is preserved. This is not difficult for either positive or negative values of the adjustment. However, negative values need a check to ensure that no nonspace characters are snipped from the start of the lines. Thus, we may as well split the functionality into two functions to perform the transformations, plus one to measure the number of leading spaces of each line and return the result as a list:def addSpaces(s, numAdd): white = " "*numAdd return white + white.join(s.splitlines(True)) def numSpaces(s): return [len(line)-len(line.lstrip( )) for line in s.splitlines( )] def delSpaces(s, numDel): if numDel > min(numSpaces(s)): raise ValueError, "removing more spaces than there are!" return '\n'.join([ line[numDel:] for line in s.splitlines( ) ]) All of these functions rely on the string method splitlines, which is similar to a split on '\n'. splitlines has the extra ability to leave the trailing newline on each line (when you call it with true as its argument). Sometimes this turns out to be handy: addSpaces could not be quite as short and sweet without this ability of the splitlines string method. Here's how we can combine these functions to build another function to delete just enough leading spaces from each line to ensure that the least-indented line of the block becomes flush left, while preserving the relative indentation of the lines: def unIndentBlock(s): return delSpaces(s, min(numSpaces(s))) See Also Library Reference and Python in a Nutshell docs on sequence types. Recipe 1.15. Expanding and Compressing Tabs Credit: Alex Martelli, David Ascher Problem You want to convert tabs in a string to the appropriate number of spaces, or vice versa. Solution Changing tabs to the appropriate number of spaces is a reasonably frequent task, easily accomplished with Python strings' expandtabs method. Because strings are immutable, the method returns a new string object, a modified copy of the original one. However, it's easy to rebind a string variable name from the original to the modified-copy value: mystring = mystring.expandtabs( ) This doesn't change the string object to which mystring originally referred, but it does rebind the name mystring to a newly created string object, a modified copy of mystring in which tabs are expanded into runs of spaces. expandtabs, by default, uses a tab length of 8; you can pass expandtabs an integer argument to use as the tab length. Changing spaces into tabs is a rare and peculiar need. Compression, if that's what you're after, is far better performed in other ways, so Python doesn't offer a built-in way to "unexpand" spaces into tabs. We can, of course, write our own function for the purpose. String processing tends to be fastest in a split/process/rejoin approach, rather than with repeated overall string transformations: def unexpand(astring, tablen=8): import re # split into alternating space and non-space sequences pieces = re.split(r'( +)', astring.expandtabs(tablen)) # keep track of the total length of the string so far lensofar = 0 for i, piece in enumerate(pieces): thislen = len(piece) lensofar += thislen if piece.isspace( ): # change each space sequences into tabs+spaces numblanks = lensofar % tablen numtabs = (thislen-numblanks+tablen-1)/tablen pieces[i] = '\t'*numtabs + ' '*numblanks return ''.join(pieces) Function unexpand, as written in this example, works only for a single-line string; to deal with a multi-line string, use ''.join([ unexpand(s) for s in astring.splitlines(True) ]). DiscussionWhile regular expressions are never indispensable for the purpose of manipulating strings in Python, they are occasionally quite handy. Function unexpand, as presented in the recipe, for example, takes advantage of one extra feature of re.split with respect to string's split method: when the regular expression contains a (parenthesized) group, re.split returns a list where the split pieces are interleaved with the "splitter" pieces. So, here, we get alternate runs of nonblanks and blanks as items of list pieces; the for loop keeps track of the length of string it has seen so far, and changes pieces that are made of blanks to as many tabs as possible, plus as many blanks are needed to maintain the overall length. Some programming tasks that could still be described as expanding tabs are unfortunately not quite as easy as just calling the expandtabs method. A category that does happen with some regularity is to fix Python source files, which use a mix of tabs and spaces for indentation (a very bad idea), so that they instead use spaces only (which is the best approach). This could entail extra complications, for example, when you need to guess the tab length (and want to end up with the standard four spaces per indentation level, which is strongly advisable). It can also happen when you need to preserve tabs that are inside strings, rather than tabs being used for indentation (because somebody erroneously used actual tabs, rather than '\t', to indicate tabs in strings), or even because you're asked to treat docstrings differently from other strings. Some cases are not too badfor example, when you want to expand tabs that occur only within runs of whitespace at the start of each line, leaving any other tab alone. A little function using a regular expression suffices: def expand_at_linestart(P, tablen=8): import re def exp(mo): return mo.group( ).expand(tablen) return ''.join([ re.sub(r'^\s+', exp, s) for s in P.splitlines(True) ]) This function expand_at_linestart exploits the re.sub function, which looks for a regular expression in a string and, each time it gets a match, calls a function, passing the match object as the argument, to obtain the string to substitute in place of the match. For convenience, expand_at_linestart is coded to deal with a multiline string argument P, performing the list comprehension over the results of the splitlines call, and the '\n'.join of the whole. Of course, this convenience does not stop the function from being able to deal with a single-line P. If your specifications regarding which tabs are to be expanded are even more complex, such as needing to deal differently with tabs depending on whether they're inside or outside of strings, and on whether or not strings are docstrings, at the very least, you need to perform a tokenization. In addition, you may also have to perform a full parse of the source code you're dealing with, rather than using simple string or regular-expression operations. If this is the case, you can expect a substantial amount of work. Some beginning pointers to help you get started may be found in Chapter 16. If you ever find yourself sweating out this kind of task, you will no doubt get excellent motivation in the future for following the normal and recommended Python style in the source code you write or edit: only spaces, four per indentation level, no tabs, and always '\t', never an actual tab character, to include a tab in a string literal. Your favorite editor can no doubt be told to enforce all of these conventions whenever a Python source file is saved; the editor that comes with IDLE (the free integrated development environment that comes with Python), for example, supports these conventions. It is much easier to arrange your editor so that the problem never arises, rather than striving to fix it after the fact! See Also Documentation for the expandtabs method of strings in the "Sequence Types" section of the Library Reference; Perl Cookbook recipe 1.7; Library Reference and Python in a Nutshell documentation of module re. Recipe 1.16. Interpolating Variables in a String Credit: Scott David Daniels Problem You need a simple way to get a copy of a string where specially marked substrings are replaced with the results of looking up the substrings in a dictionary. Solution Here is a solution that works in Python 2.3 as well as in 2.4: def expand(format, d, marker='"', safe=False): if safe: def lookup(w): return d.get(w, w.join(marker*2)) else: def lookup(w): return d[w] parts = format.split(marker) parts[1::2] = map(lookup, parts[1::2]) return ''.join(parts) if _ _name_ _ == '_ _main_ _': print expand('just "a" test', {'a': 'one'}) # emits: just one test When the parameter safe is False, the default, every marked substring must be found in dictionary d, otherwise expand terminates with a KeyError exception. When parameter safe is explicitly passed as true, marked substrings that are not found in the dictionary are just left intact in the output string. Discussion The code in the body of the expand function has some points of interest. It defines one of two different nested functions (with the name of lookup either way), depending on whether the expansion is required to be safe. Safe means no KeyError exception gets raised for marked strings not found in the dictionary. If not required to be safe (the default), lookup just indexes into dictionary d and raises an error if the substring is not found. But, if lookup is required to be "safe", it uses d's method get and supplies as the default the substring being looked up, with a marker on either side. In this way, by passing safe as true, you may choose to have unknown formatting markers come right through to the output rather than raising exceptions. marker+w+marker would be an obvious alternative to the chosen w.join(marker*2), but I've chosen the latter exactly to display a non-obvious but interesting way to construct such a quoted string. With either version of lookup, expand operates according to the split/modify/join idiom that is so important for Python string processing. The modify part, in expand's case, makes use of the possibility of accessing and modifying a list's slice with a "step" or "stride". Specifically, expand accesses and rebinds all of those items of parts that lie at an odd index, because those items are exactly the ones that were enclosed between a pair of markers in the original format string. Therefore, they are the marked substrings that may be looked up in the dictionary.The syntax of format strings accepted by this recipe's function expand is more flexible than the $-based syntax of string.Template. You can specify a different marker when you want your format string to contain double quotes, for example. There is no constraint for each specially marked substring to be an identifier, so you can easily interpolate Python expressions (with a d whose _ _getitem_ _ performs an eval) or any other kind of placeholder. Moreover, you can easily get slightly different, useful effects. For example: print expand('just "a" ""little"" test', {'a' : 'one', '' : '"'}) emits just one "little" test. Advanced users can customize Python 2.4's string.Template class, by inheritance, to match all of these capabilities, and more, but this recipe's little expand function is still simpler to use in some flexible ways. See Also Library Reference docs for string.Template (Python 2.4, only), the section on sequence types (for string methods split and join, and for slicing operations), and the section on dictionaries (for indexing and the get method). For more information on Python 2.4's string.Template class, see Recipe 1.17. Recipe 1.17. Interpolating Variables in a Stringin Python 2.4 Credit: John Nielsen, Lawrence Oluyede, Nick Coghlan Problem Using Python 2.4, you need a simple way to get a copy of a string where specially marked identifiers are replaced with the results of looking up the identifiers in a dictionary. Solution Python 2.4 offers the new string.Template class for this purpose. Here is a snippet of code showing how to use that class: import string # make a template from a string where some identifiers are marked with $ new_style = string.Template('this is $thing') # use the substitute method of the template with a dictionary argument: print new_style.substitute({'thing':5}) # emits: this is 5 print new_style.substitute({'thing':'test'}) # emits: this is test # alternatively, you can pass keyword-arguments to 'substitute': print new_style.substitute(thing=5) # emits: this is 5 print new_style.substitute(thing='test') # emits: this is test Discussion In Python 2.3, a format string for identifier-substitution has to be expressed in a less simple format: old_style = 'this is %(thing)s' with the identifier in parentheses after a %, and an s right after the closed parenthesis. Then, you use the % operator, with the format string on the left of the operator, and a dictionary on the right: print old_style % {'thing':5} # emits: this is 5 print old_style % {'thing':'test'} # emits: this is test Of course, this code keeps working in Python 2.4, too. However, the new string.Template class offers a simpler alternative. When you build a string.Template instance, you may include a dollar sign ($) by doubling it, and you may have the interpolated identifier immediately followed by letters or digits by enclosing it in curly braces ({ }). Here is an example that requires both of these refinements: form_letter = '''Dear $customer, I hope you are having a great time.If you do not find Room $room to your satisfaction, let us know. Please accept this $$5 coupon. Sincerely, $manager ${name}Inn''' letter_template = string.Template(form_letter) print letter_template.substitute({'name':'Sleepy', 'customer':'Fred Smith', 'manager':'Barney Mills', 'room':307, }) This snippet emits the following output: Dear Fred Smith, I hope you are having a great time. If you do not find Room 307 to your satisfaction, let us know. Please accept this $5 coupon. Sincerely, Barney Mills SleepyInn Sometimes, the handiest way to prepare a dictionary to be used as the argument to the substitute method is to set local variables, and then pass as the argument locals( ) (the artificial dictionary whose keys are the local variables, each with its value associated): msg = string.Template('the square of $number is $square') for number in range(10): square = number * number print msg.substitute(locals( )) Another handy alternative is to pass the values to substitute using keyword argument syntax rather than a dictionary: msg = string.Template('the square of $number is $square') for i in range(10): print msg.substitute(number=i, square=i*i) You can even pass both a dictionary and keyword arguments: msg = string.Template('the square of $number is $square') for number in range(10): print msg.substitute(locals( ), square=number*number) In case of any conflict between entries in the dictionary and the values explicitly passed as keyword arguments, the keyword arguments take precedence. For example: msg = string.Template('an $adj $msg') adj = 'interesting' print msg.substitute(locals( ), msg='message') # emits an interesting message See Also Library Reference docs for string.Template (2.4 only) and the locals built-in function. Recipe 1.18. Replacing Multiple Patterns in a Single Pass Credit: Xavier Defrang, Alex Martelli Problem You need to perform several string substitutions on a string. Solution Sometimes regular expressions afford the fastest solution even in cases where their applicability is not obvious. The powerful sub method of re objects (from the re module in the standard library) makes regular expressions particularly good at performing string substitutions. Here is a function returning a modified copy of an input string, where each occurrence of any string that's a key in a given dictionary is replaced by the corresponding value in the dictionary: import re def multiple_replace(text, adict): rx = re.compile('|'.join(map(re.escape, adict))) def one_xlat(match): return adict[match.group(0)] return rx.sub(one_xlat, text) Discussion This recipe shows how to use the Python standard re module to perform single-pass multiple- string substitution using a dictionary. Let's say you have a dictionary-based mapping between strings. The keys are the set of strings you want to replace, and the corresponding values are the strings with which to replace them. You could perform the substitution by calling the string method replace for each key/value pair in the dictionary, thus processing and creating a new copy of the entire text several times, but it is clearly better and faster to do all the changes in a single pass, processing and creating a copy of the text only once. re.sub's callback facility makes this better approach quite easy. First, we have to build a regular expression from the set of keys we want to match. Such a regular expression has a pattern of the form a1|a2|...|aN, made up of the N strings to be substituted, joined by vertical bars, and it can easily be generated using a one-liner, as shown in the recipe. Then, instead of giving re.sub a replacement string, we pass it a callback argument. re.sub then calls this object for each match, with a re.MatchObject instance as its only argument, and it expects the replacement string for that match as the call's result. In our case, the callback just has to look up the matched text in the dictionary and return the corresponding value. The function multiple_replace presented in the recipe recomputes the regular expression and redefines the one_xlat auxiliary function each time you call it. Often, you must perform substitutions on multiple strings based on the same, unchanging translation dictionary and would prefer to pay these setup prices only once. For such needs, you may prefer the following closure- based approach:import re def make_xlat(*args, **kwds): adict = dict(*args, **kwds) rx = re.compile('|'.join(map(re.escape, adict))) def one_xlat(match): return adict[match.group(0)] def xlat(text): return rx.sub(one_xlat, text) return xlat You can call make_xlat, passing as its argument a dictionary, or any other combination of arguments you could pass to built-in dict in order to construct a dictionary; make_xlat returns a xlat closure that takes as its only argument text the string on which the substitutions are desired and returns a copy of text with all the substitutions performed. Here's a usage example for each half of this recipe. We would normally have such an example as a part of the same .py source file as the functions in the recipe, so it is guarded by the traditional Python idiom that runs it if and only if the module is called as a main script: if _ _name_ _ == "_ _main_ _": text = "Larry Wall is the creator of Perl" adict = { "Larry Wall" : "Guido van Rossum", "creator" : "Benevolent Dictator for Life", "Perl" : "Python", } print multiple_replace(text, adict) translate = make_xlat(adict) print translate(text) Substitutions such as those performed by this recipe are often intended to operate on entire words, rather than on arbitrary substrings. Regular expressions are good at picking up the beginnings and endings of words, thanks to the special sequence r'\b'. We can easily make customized versions of either multiple_replace or make_xlat by simply changing the one line in which each of them builds and assigns the regular expression object rx into a slightly different form: rx = re.compile(r'\b%s\b' % r'\b|\b'.join(map(re.escape, adict))) The rest of the code is just the same as shown earlier in this recipe. However, this sameness is not necessarily good news: it suggests that if we need many similarly customized versions, each building the regular expression in slightly different ways, we'll end up doing a lot of copy-and- paste coding, which is the worst form of code reuse, likely to lead to high maintenance costs in the future. A key rule of good coding is: "once, and only once!" When we notice that we are duplicating code, we should notice this symptom as a "code smell," and refactor our code for better reuse. In this case, for ease of customization, we need a class rather than a function or closure. For example, here's how to write a class that works very similarly to make_xlat but can be customized by subclassing and overriding: class make_xlat: def _ _init_ _(self, *args, **kwds): self.adict = dict(*args, **kwds) self.rx = self.make_rx( ) def make_rx(self): return re.compile('|'.join(map(re.escape, self.adict))) def one_xlat(self, match): return self.adict[match.group(0)] def _ _call_ _(self, text): return self.rx.sub(self.one_xlat, text) This is a "drop-in replacement" for the function of the same name: in other words, a snippet such as the one we showed, with the if _ _name_ _ == '_ _main_ _' guard, works identically when make_xlat is this class rather than the previously shown function. The function is simpler and faster, but the class' important advantage is that it can easily be customized in the usual object- oriented waysubclassing it, and overriding some method. To translate by whole words, for example, all we need to code is: class make_xlat_by_whole_words(make_xlat): def make_rx(self): return re.compile(r'\b%s\b' % r'\b|\b'.join(map(re.escape, self.adict))) Ease of customization by subclassing and overriding helps you avoid copy-and-paste coding, and this is sometimes an excellent reason to prefer object-oriented structures over simpler functional structures, such as closures. Of course, just because some functionality is packaged as a class doesn't magically make it customizable in just the way you want. Customizability also requires some foresight in dividing the functionality into separately overridable methods that correspond to the right pieces of overall functionality. Fortunately, you don't have to get it right the first time; when code does not have the optimal internal structure for the task at hand (in this specific example, for reuse by subclassing and selective overriding), you can and should refactor the code so that its internal structure serves your needs. Just make sure you have a suitable battery of tests ready to run to ensure that your refactoring hasn't broken anything, and then you can refactor to your heart's content. See http://www.refactoring.com for more information on the important art and practice of refactoring. See Also Documentation for the re module in the Library Reference and Python in a Nutshell; the Refactoring home page (http://www.refactoring.com). Recipe 1.19. Checking a String for Any of Multiple Endings Credit: Michele Simionato Problem For a certain string s, you must check whether s has any of several endings; in other words, you need a handy, elegant equivalent of s.endswith(end1) or s.endswith(end2) or s.endswith(end3) and so on. Solution The itertools.imap function is just as handy for this task as for many of a similar nature: import itertools def anyTrue(predicate, sequence): return True in itertools.imap(predicate, sequence) def endsWith(s, *endings): return anyTrue(s.endswith, endings) Discussion A typical use for endsWith might be to print all names of image files in the current directory: import os for filename in os.listdir('.'): if endsWith(filename, '.jpg', '.jpeg', '.gif'): print filename The same general idea shown in this recipe's Solution is easily applied to other tasks related to checking a string for any of several possibilities. The auxiliary function anyTrue is general and fast, and you can pass it as its first argument (the predicate) other bound methods, such as s.startswith or s._ _contains_ _. Indeed, perhaps it would be better to do without the helper function endsWithafter all, directly coding if anyTrue(filename.endswith, (".jpg", ".gif", ".png")): seems to be already readable enough.Bound Methods Whenever a Python object supplies a method, you can get the method, already bound to the object, by just accessing the method on the object. (For example, you can assign it, pass it as an argument, return it as a function's result, etc.) For example: L = ['fee', 'fie', 'foo'] x = L.append Now, name x refers to a bound method of list object L. Calling, say, x('fum') is the same as calling L.append('fum'): either call mutates object L into ['fee', 'fie', 'foo', 'fum']. If you access a method on a type or class, rather than an instance of the type or class, you get an unbound method, not "attached" to any particular instance of the type or class: when you call it, you need to pass as its first argument an instance of that type or class. For example, if you set y = list.append, you cannot just call y('I')Python couldn't possibly guess which list you want to append I to! You can, however, call y(L, 'I'), and that is just the same as calling L.append('I') (as long as isinstance(L, list)). This recipe originates from a discussion on news:comp.lang.python. and summarizes inputs from many people, including Raymond Hettinger, Chris Perkins, Bengt Richter and others. See Also Library Reference and Python in a Nutshell docs for itertools and string methods. Recipe 1.20. Handling International Text with Unicode Credit: Holger Krekel Problem You need to deal with text strings that include non-ASCII characters. Solution Python has a first class unicode type that you can use in place of the plain bytestring str type. It's easy, once you accept the need to explicitly convert between a bytestring and a Unicode string: >>> german_ae = unicode('\xc3\xa4', 'utf8') Here german_ae is a unicode string representing the German lowercase a with umlaut (i.e., diaeresis) character "ae". It has been constructed from interpreting the bytestring '\xc3\xa4' according to the specified UTF-8 encoding. There are many encodings, but UTF-8 is often used because it is universal (UTF-8 can encode any Unicode string) and yet fully compatible with the 7-bit ASCII set (any ASCII bytestring is a correct UTF-8-encoded string). Once you cross this barrier, life is easy! You can manipulate this Unicode string in practically the same way as a plain str string: >>> sentence = "This is a " + german_ae >>> sentence2 = "Easy!" >>> para = ". ".join([sentence, sentence2]) Note that para is a Unicode string, because operations between a unicode string and a bytestring always result in a unicode stringunless they fail and raise an exception: >>> bytestring = '\xc3\xa4' # Uuh, some non-ASCII bytestring! >>> german_ae += bytestring UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 0: ordinal not in range(128) The byte '0xc3' is not a valid character in the 7-bit ASCII encoding, and Python refuses to guess an encoding. So, being explicit about encodings is the crucial point for successfully using Unicode strings with Python. Discussion Unicode is easy to handle in Python, if you respect a few guidelines and learn to deal with common problems. This is not to say that an efficient implementation of Unicode is an easy task. Luckily, as with other hard problems, you don't have to care much: you can just use the efficient implementation of Unicode that Python provides.The most important issue is to fully accept the distinction between a bytestring and a unicode string. As exemplified in this recipe's solution, you often need to explicitly construct a unicode string by providing a bytestring and an encoding. Without an encoding, a bytestring is basically meaningless, unless you happen to be lucky and can just assume that the bytestring is text in ASCII. The most common problem with using Unicode in Python arises when you are doing some text manipulation where only some of your strings are unicode objects and others are bytestrings. Python makes a shallow attempt to implicitly convert your bytestrings to Unicode. It usually assumes an ASCII encoding, though, which gives you UnicodeDecodeError exceptions if you actually have non-ASCII bytes somewhere. UnicodeDecodeError tells you that you mixed Unicode and bytestrings in such a way that Python cannot (doesn't even try to) guess the text your bytestring might represent. Developers from many big Python projects have come up with simple rules of thumb to prevent such runtime UnicodeDecodeErrors, and the rules may be summarized into one sentence: always do the conversion at IO barriers. To express this same concept a bit more extensively: Whenever your program receives text data "from the outside" (from the network, from a file, from user input, etc.), construct unicode objects immediately. Find out the appropriate encoding, for example, from an HTTP header, or look for an appropriate convention to determine the encoding to use. Whenever your program sends text data "to the outside" (to the network, to some file, to the user, etc.), determine the correct encoding, and convert your text to a bytestring with that encoding. (Otherwise, Python attempts to convert Unicode to an ASCII bytestring, likely producing UnicodeEncodeErrors, which are just the converse of the UnicodeDecodeErrors previously mentioned). With these two rules, you will solve most Unicode problems. If you still get UnicodeErrors of either kind, look for where you forgot to properly construct a unicode object, forgot to properly convert back to an encoded bytestring, or ended up using an inappropriate encoding due to some mistake. (It is quite possible that such encoding mistakes are due to the user, or some other program that is interacting with yours, not following the proper encoding rules or conventions.) In order to convert a Unicode string back to an encoded bytestring, you usually do something like: >>> bytestring = german_ae.decode('latin1') >>> bytestring '\xe4' Now bytestring is a German ae character in the 'latin1' encoding. Note how '\xe4' (in Latin1) and the previously shown '\xc3\xa4' (in UTF-8) represent the same German character, but in different encodings. By now, you can probably imagine why Python refuses to guess among the hundreds of possible encodings. It's a crucial design choice, based on one of the Zen of Python principles: "In the face of ambiguity, resist the temptation to guess." At any interactive Python shell prompt, enter the statement import this to read all of the important principles that make up the Zen of Python. See Also Unicode is a huge topic, but a recommended book is Unicode: A Primer, by Tony Graham (Hungry Minds, Inc.)details are available at http://www.menteith.com/unicode/primer/; and a short but complete article from Joel Spolsky, "The Absolute Minimum Every Software Developer Absolutely, Positively Must Know About Unicode and Character Sets (No Excuses)!," located athttp://www.joelonsoftware.com/articles/Unicode.html. See also the Library Reference and Python in a Nutshell documentation about the built-in str and unicode types and modules unidata and codecs; also, Recipe 1.21 and Recipe 1.22. Recipe 1.21. Converting Between Unicode and Plain Strings Credit: David Ascher, Paul Prescod Problem You need to deal with textual data that doesn't necessarily fit in the ASCII character set. Solution Unicode strings can be encoded in plain strings in a variety of ways, according to whichever encoding you choose: unicodestring = u"Hello world" # Convert Unicode to plain Python string: "encode" utf8string = unicodestring.encode("utf-8") asciistring = unicodestring.encode("ascii") isostring = unicodestring.encode("ISO-8859-1") utf16string = unicodestring.encode("utf-16") # Convert plain Python string to Unicode: "decode" plainstring1 = unicode(utf8string, "utf-8") plainstring2 = unicode(asciistring, "ascii") plainstring3 = unicode(isostring, "ISO-8859-1") plainstring4 = unicode(utf16string, "utf-16") assert plainstring1 == plainstring2 == plainstring3 == plainstring4 Discussion If you find yourself dealing with text that contains non-ASCII characters, you have to learn about Unicodewhat it is, how it works, and how Python uses it. The preceding Recipe 1.20 offers minimal but crucial practical tips, and this recipe tries to offer more perspective. You don't need to know everything about Unicode to be able to solve real-world problems with it, but a few basic tidbits of knowledge are indispensable. First, you must understand the difference between bytes and characters. In older, ASCII-centric languages and environments, bytes and characters are treated as if they were the same thing. A byte can hold up to 256 different values, so these environments are limited to dealing with no more than 256 distinct characters. Unicode, on the other hand, has tens of thousands of characters, which means that each Unicode character takes more than one byte; thus you need to make the distinction between characters and bytes. Standard Python strings are really bytestrings, and a Python character, being such a string of length 1, is really a byte. Other terms for an instance of the standard Python string type are 8-bit string and plain string. In this recipe we call such instances bytestrings, to remind you of their byte orientation. A Python Unicode character is an abstract object big enough to hold any character, analogous to Python's long integers. You don't have to worry about the internal representation; the representation of Unicode characters becomes an issue only when you are trying to send them tosome byte-oriented function, such as the write method of files or the send method of network sockets. At that point, you must choose how to represent the characters as bytes. Converting from Unicode to a bytestring is called encoding the string. Similarly, when you load Unicode strings from a file, socket, or other byte-oriented object, you need to decode the strings from bytes to characters. Converting Unicode objects to bytestrings can be achieved in many ways, each of which is called an encoding. For a variety of historical, political, and technical reasons, there is no one "right" encoding. Every encoding has a case-insensitive name, and that name is passed to the encode and decode methods as a parameter. Here are a few encodings you should know about: The UTF-8 encoding can handle any Unicode character. It is also backwards compatible with ASCII, so that a pure ASCII file can also be considered a UTF-8 file, and a UTF-8 file that happens to use only ASCII characters is identical to an ASCII file with the same characters. This property makes UTF-8 very backwards-compatible, especially with older Unix tools. UTF-8 is by far the dominant encoding on Unix, as well as the default encoding for XML documents. UTF-8's primary weakness is that it is fairly inefficient for eastern-language texts. The UTF-16 encoding is favored by Microsoft operating systems and the Java environment. It is less efficient for western languages but more efficient for eastern ones. A variant of UTF-16 is sometimes known as UCS-2. The ISO-8859 series of encodings are supersets of ASCII, each able to deal with 256 distinct characters. These encodings cannot support all of the Unicode characters; they support only some particular language or family of languages. ISO-8859-1, also known as "Latin-1", covers most western European and African languages, but not Arabic. ISO-8859- 2, also known as "Latin-2", covers many eastern European languages such as Hungarian and Polish. ISO-8859-15, very popular in Europe these days, is basically the same as ISO- 8859-1 with the addition of the Euro currency symbol as a character. If you want to be able to encode all Unicode characters, you'll probably want to use UTF-8. You will need to deal with the other encodings only when you are handed data in those encodings created by some other application or input device, or vice versa, when you need to prepare data in a specified encoding to accommodate another application downstream of yours, or an output device. In particular, Recipe 1.22 shows how to handle the case in which the downstream application or device is driven from your program's standard output stream. See Also Unicode is a huge topic, but a recommended book is Tony Graham, Unicode: A Primer (Hungry Minds)details are available at http://www.menteith.com/unicode/primer/; and a short, but complete article from Joel Spolsky, "The Absolute Minimum Every Software Developer Absolutely, Positively Must Know About Unicode and Character Sets (No Excuses)!" is located at http://www.joelonsoftware.com/articles/Unicode.html. See also the Library Reference and Python in a Nutshell documentation about the built-in str and unicode types, and modules unidata and codecs; also, Recipe 1.20 and Recipe 1.22. Recipe 1.22. Printing Unicode Charactersto Standard Output Credit: David Ascher Problem You want to print Unicode strings to standard output (e.g., for debugging), but they don't fit in the default encoding. Solution Wrap the sys.stdout stream with a converter, using the codecs module of Python's standard library. For example, if you know your output is going to a terminal that displays characters according to the ISO-8859-1 encoding, you can code: import codecs, sys sys.stdout = codecs.lookup('iso8859-1')[-1](sys.stdout) Discussion Unicode strings live in a large space, big enough for all of the characters in every language worldwide, but thankfully the internal representation of Unicode strings is irrelevant for users of Unicode. Alas, a file stream, such as sys.stdout, deals with bytes and has an encoding associated with it. You can change the default encoding that is used for new files by modifying the site module. That, however, requires changing your entire Python installation, which is likely to confuse other applications that may expect the encoding you originally configured Python to use (typically the Python standard encoding, which is ASCII). Therefore, this kind of modification is not to be recommended. This recipe takes a sounder approach: it rebinds sys.stdout as a stream that expects Unicode input and outputs it in ISO-8859-1 (also known as "Latin-1"). This approach doesn't change the encoding of any previous references to sys.stdout, as illustrated here. First, we keep a reference to the original, ASCII-encoded sys.stdout: >>> old = sys.stdout Then, we create a Unicode string that wouldn't normally be able to go through sys.stdout: >>> char = u"\N{LATIN SMALL LETTER A WITH DIAERESIS}" >>> print char Traceback (most recent call last): File "", line 1, in ? UnicodeError: ASCII encoding error: ordinal not in range(128) If you don't get an error from this operation, it's because Python thinks it knows which encoding your "terminal" is using (in particular, Python is likely to use the right encoding if your "terminal" is IDLE, the free development environment that comes with Python). But, suppose you do getthis error, or get no error but the output is not the character you expected, because your "terminal" uses UTF-8 encoding and Python does not know about it. When that is the case, we can just wrap sys.stdout in the codecs stream writer for UTF-8, which is a much richer encoding, then rebind sys.stdout to it and try again: >>> sys.stdout = codecs.lookup('utf-8')[-1](sys.stdout) >>> print char ä This approach works only if your "terminal", terminal emulator, or other window in which you're running the interactive Python interpreter supports the UTF-8 encoding, with a font rich enough to display all the characters you need to output. If you don't have such a program or device available, you may be able to find a suitable one for your platform in the form of a free program downloadable from the Internet. Python tries to determine which encoding your "terminal" is using and sets that encoding's name as attribute sys.stdout.encoding. Sometimes (alas, not always) it even manages to get it right. IDLE already wraps your sys.stdout, as suggested in this recipe, so, within the environment's interactive Python shell, you can directly print Unicode strings. See Also Documentation for the codecs and site modules, and setdefaultencoding in module sys, in the Library Reference and Python in a Nutshell; Recipe 1.20 and Recipe 1.21. Recipe 1.23. Encoding Unicode Data for XML and HTML Credit: David Goodger, Peter Cogolo Problem You want to encode Unicode text for output in HTML, or some other XML application, using a limited but popular encoding such as ASCII or Latin-1. Solution Python provides an encoding error handler named xmlcharrefreplace, which replaces all characters outside of the chosen encoding with XML numeric character references: def encode_for_xml(unicode_data, encoding='ascii'): return unicode_data.encode(encoding, 'xmlcharrefreplace') You could use this approach for HTML output, too, but you might prefer to use HTML's symbolic entity references instead. For this purpose, you need to define and register a customized encoding error handler. Implementing that handler is made easier by the fact that the Python Standard Library includes a module named htmlentitydefs that holds HTML entity definitions: import codecs from htmlentitydefs import codepoint2name def html_replace(exc): if isinstance(exc, (UnicodeEncodeError, UnicodeTranslateError)): s = [ u'&%s;' % codepoint2name[ord(c)] for c in exc.object[exc.start:exc.end] ] return ''.join(s), exc.end else: raise TypeError("can't handle %s" % exc._ _name_ _) codecs.register_error('html_replace', html_replace) After registering this error handler, you can optionally write a function to wrap its use: def encode_for_html(unicode_data, encoding='ascii'): return unicode_data.encode(encoding, 'html_replace') Discussion As with any good Python module, this module would normally proceed with an example of its use, guarded by an if _ _name_ _ == '_ _main_ _' test: if _ _name_ _ == '_ _main_ _': # demo data = u'''\ Encoding Test

accented characters:

  • \xe0 (a + grave)
  • \xe7 (c + cedilla)
  • \xe9 (e + acute)

symbols:

  • \xa3 (British pound)
  • \u20ac (Euro)
  • \u221e (infinity)
''' print encode_for_xml(data) print encode_for_html(data) If you run this module as a main script, you will then see such output as (from function encode_for_xml):
  • à (a + grave)
  • ç (c + cedilla)
  • é (e + acute) ...
  • £ (British pound)
  • € (Euro)
  • ∞ (infinity) as well as (from function encode_for_html):
  • à (a + grave)
  • ç (c + cedilla)
  • é (e + acute) ...
  • £ (British pound)
  • € (Euro)
  • ∞ (infinity) There is clearly a niche for each case, since encode_for_xml is more general (you can use it for any XML application, not just HTML), but encode_for_html may produce output that's easier to readshould you ever need to look at it directly, edit it further, and so on. If you feed either form to a browser, you should view it in exactly the same way. To visualize both forms of encoding in a browser, run this recipe's module as a main script, redirect the output to a disk file, and use a text editor to separate the two halves before you view them with a browser. (Alternatively, run the script twice, once commenting out the call to encode_for_xml, and once commenting out the call to encode_for_html.) Remember that Unicode data must always be encoded before being printed or written out to a file. UTF-8 is an ideal encoding, since it can handle any Unicode character. But for many users and applications, ASCII or Latin-1 encodings are often preferred over UTF-8. When the Unicode data contains characters that are outside of the given encoding (e.g., accented characters and most symbols are not encodable in ASCII, and the "infinity" symbol is not encodable in Latin-1), these encodings cannot handle the data on their own. Python supports a built-in encoding error handler called xmlcharrefreplace, which replaces unencodable characters with XML numeric character references, such as ∞ for the "infinity" symbol. This recipe shows how to writeand register another similar error handler, html_replace, specifically for producing HTML output. html_replace replaces unencodable characters with more readable HTML symbolic entity references, such as ∞ for the "infinity" symbol. html_replace is less general than xmlcharrefreplace, since it does not support all Unicode characters and cannot be used with non-HTML applications; however, it can still be useful if you want HTML output that is as readable as possible in a "view page source" context. Neither of these error handlers makes sense for output that is neither HTML nor some other form of XML. For example, TeX and other markup languages do not recognize XML numeric character references. However, if you know how to build an arbitrary character reference for such a markup language, you may modify the example error handler html_replace shown in this recipe's Solution to code and register your own encoding error handler. An alternative (and very effective!) way to perform encoding of Unicode data into a file, with a given encoding and error handler of your choice, is offered by the codecs module in Python's standard library: outfile = codecs.open('out.html', mode='w', encoding='ascii', errors='html_replace') You can now use outfile.write(unicode_data) for any arbitrary Unicode string unicode_data, and all the encoding and error handling will be taken care of transparently. When your output is finished, of course, you should call outfile.close( ). See Also Library Reference and Python in a Nutshell docs for modules codecs and htmlentitydefs. Recipe 1.24. Making Some Strings Case-Insensitive Credit: Dale Strickland-Clark, Peter Cogolo, Mark McMahon Problem You want to treat some strings so that all comparisons and lookups are case-insensitive, while all other uses of the strings preserve the original case. Solution The best solution is to wrap the specific strings in question into a suitable subclass of str: class iStr(str): """ Case insensitive string class. Behaves just like str, except that all comparisons and lookups are case insensitive. """ def _ _init_ _(self, *args): self._lowered = str.lower(self) def _ _repr_ _(self): return '%s(%s)' % (type(self)._ _name_ _, str._ _repr_ _(self)) def _ _hash_ _(self): return hash(self._lowered) def lower(self): return self._lowered def _make_case_insensitive(name): ''' wrap one method of str into an iStr one, case-insensitive ''' str_meth = getattr(str, name) def x(self, other, *args): ''' try lowercasing 'other', which is typically a string, but be prepared to use it as-is if lowering gives problems, since strings CAN be correctly compared with non-strings. ''' try: other = other.lower( ) except (TypeError, AttributeError, ValueError): pass return str_meth(self._lowered, other, *args) # in Python 2.4, only, add the statement: x.func_name = name setattr(iStr, name, x) # apply the _make_case_insensitive function to specified methods for name in 'eq lt le gt gt ne cmp contains'.split( ): _make_case_insensitive('_ _%s_ _' % name) for name in 'count endswith find index rfind rindex startswith'.split( ): _make_case_insensitive(name) # note that we don't modify methods 'replace', 'split', 'strip', ... # of course, you can add modifications to them, too, if you prefer that. del _make_case_insensitive # remove helper function, not needed any more DiscussionSome implementation choices in class iStr are worthy of notice. First, we choose to generate the lowercase version once and for all, in method _ _init_ _, since we envision that in typical uses of iStr instances, this version will be required repeatedly. We hold that version in an attribute that is private, but not overly so (i.e., has a name that begins with one underscore, not two), because if iStr gets subclassed (e.g., to make a more extensive version that also offers case- insensitive splitting, replacing, etc., as the comment in the "Solution" suggests), iStr's subclasses are quite likely to want to access this crucial "implementation detail" of superclass iStr! We do not offer "case-insensitive" versions of such methods as replace, because it's anything but clear what kind of input-output relation we might want to establish in the general case. Application-specific subclasses may therefore be the way to provide this functionality in ways appropriate to a given application. For example, since the replace method is not wrapped, calling replace on an instance of iStr returns an instance of str, not of iStr. If that is a problem in your application, you may want to wrap all iStr methods that return strings, simply to ensure that the results are made into instances of iStr. For that purpose, you need another, separate helper function, similar but not identical to the _make_case_insensitive one shown in the "Solution": def _make_return_iStr(name): str_meth = getattr(str, name) def x(*args): return iStr(str_meth(*args)) setattr(iStr, name, x) and you need to call this helper function _make_return_iStr on all the names of relevant string methods returning strings such as: for name in 'center ljust rjust strip lstrip rstrip'.split( ): _make_return_iStr(name) Strings have about 20 methods (including special methods such as _ _add_ _ and _ _mul_ _) that you should consider wrapping in this way. You can also wrap in this way some additional methods, such as split and join, which may require special handling, and others, such as encode and decode, that you cannot deal with unless you also define a case-insensitive unicode subtype. In practice, one can hope that not every single one of these methods will prove problematic in a typical application. However, as you can see, the very functional richness of Python strings makes it a bit of work to customize string subtypes fully, in a general way without depending on the needs of a specific application. The implementation of iStr is careful to avoid the boilerplate code (meaning repetitious and therefore bug-prone code) that we'd need if we just overrode each needed method of str in the normal way, with def statements in the class body. A custom metaclass or other such advanced technique would offer no special advantage in this case, so the boilerplate avoidance is simply obtained with one helper function that generates and installs wrapper closures, and two loops using that function, one for normal methods and one for special ones. The loops need to be placed after the class statement, as we do in this recipe's Solution, because they need to modify the class object iStr, and the class object doesn't exist yet (and thus cannot be modified) until the class statement has completed. In Python 2.4, you can reassign the func_name attribute of a function object, and in this case, you should do so to get clearer and more readable results when introspection (e.g., the help function in an interactive interpreter session) is applied to an iStr instance. However, Python 2.3 considers attribute func_name of function objects to be read-only; therefore, in this recipe's Solution, we have indicated this possibility only in a comment, to avoid losing Python 2.3 compatibility over such a minor issue. Case-insensitive (but case-preserving) strings have many uses, from more tolerant parsing of user input, to filename matching on filesystems that share this characteristic, such as all of Windows filesystems and the Macintosh default filesystem. You might easily find yourself creatinga variety of "case-insensitive" container types, such as dictionaries, lists, sets, and so onmeaning containers that go out of their way to treat string-valued keys or items as if they were case- insensitive. Clearly a better architecture is to factor out the functionality of "case-insensitive" comparisons and lookups once and for all; with this recipe in your toolbox, you can just add the required wrapping of strings into iStr instances wherever you may need it, including those times when you're making case-insensitive container types. For example, a list whose items are basically strings, but are to be treated case-insensitively (for sorting purposes and in such methods as count and index), is reasonably easy to build on top of iStr: class iList(list): def _ _init_ _(self, *args): list._ _init_ _(self, *args) # rely on _ _setitem_ _ to wrap each item into iStr... self[:] = self wrap_each_item = iStr def _ _setitem_ _(self, i, v): if isinstance(i, slice): v = map(self.wrap_each_item, v) else: v = self.wrap_each_item(v) list._ _setitem_ _(self, i, v) def append(self, item): list.append(self, self.wrap_each_item(item)) def extend(self, seq): list.extend(self, map(self.wrap_each_item, seq)) Essentially, all we're doing is ensuring that every item that gets into an instance of iList gets wrapped by a call to iStr, and everything else takes care of itself. Incidentally, this example class iList is accurately coded so that you can easily make customized subclasses of iList to accommodate application-specific subclasses of iStr: all such a customized subclass of iList needs to do is override the single class-level member named wrap_each_item. See Also Library Reference and Python in a Nutshell sections on str, string methods, and special methods used in comparisons and hashing. Recipe 1.25. Converting HTML Documents to Texton a Unix Terminal Credit: Brent Burley, Mark Moraes Problem You need to visualize HTML documents as text, with support for bold and underlined display on your Unix terminal. Solution The simplest approach is to code a filter script, taking HTML on standard input and emitting text and terminal control sequences on standard output. Since this recipe only targets Unix, we can get the needed terminal control sequences from the "Unix" command tput, via the function popen of the Python Standard Library module os: #!/usr/bin/env python import sys, os, htmllib, formatter # use Unix tput to get the escape sequences for bold, underline, reset set_bold = os.popen('tput bold').read( ) set_underline = os.popen('tput smul').read( ) perform_reset = os.popen('tput sgr0').read( ) class TtyFormatter(formatter.AbstractFormatter): ''' a formatter that keeps track of bold and italic font states, and emits terminal control sequences accordingly. ''' def _ _init_ _(self, writer): # first, as usual, initialize the superclass formatter.AbstractFormatter._ _init_ _(self, writer) # start with neither bold nor italic, and no saved font state self.fontState = False, False self.fontStack = [ ] def push_font(self, font): # the `font' tuple has four items, we only track the two flags # about whether italic and bold are active or not size, is_italic, is_bold, is_tt = font self.fontStack.append((is_italic, is_bold)) self._updateFontState( ) def pop_font(self, *args): # go back to previous font state try: self.fontStack.pop( ) except IndexError: pass self._updateFontState( ) def updateFontState(self): # emit appropriate terminal control sequences if the state of # bold and/or italic(==underline) has just changed try: newState = self.fontStack[-1] except IndexError: newState = False, False if self.fontState != newState: # relevant state change: reset terminal print perform_reset, # set underine and/or bold if needed if newState[0]: print set_underline, if newState[1]: print set_bold, # remember the two flags as our current font-state self.fontState = newState # make writer, formatter and parser objects, connecting them as needed myWriter = formatter.DumbWriter( ) if sys.stdout.isatty( ): myFormatter = TtyFormatter(myWriter) else: myFormatter = formatter.AbstractFormatter(myWriter) myParser = htmllib.HTMLParser(myFormatter) # feed all of standard input to the parser, then terminate operations myParser.feed(sys.stdin.read( )) myParser.close( ) Discussion The basic formatter.AbstractFormatter class, offered by the Python Standard Library, should work just about anywhere. On the other hand, the refinements in the TtyFormatter subclass that's the focus of this recipe depend on using a Unix-like terminal, and more specifically on the availability of the tput Unix command to obtain information on the escape sequences used to get bold or underlined output and to reset the terminal to its base state. Many systems that do not have Unix certification, such as Linux and Mac OS X, do have a perfectly workable tput command and therefore can use this recipe's TtyFormatter subclass just fine. In other words, you can take the use of the word "Unix" in this recipe just as loosely as you can take it in just about every normal discussion: take it as meaning "*ix," if you will. If your "terminal" emulator supports other escape sequences for controlling output appearance, you should be able to adapt this TtyFormatter class accordingly. For example, on Windows, a cmd.exe command window should, I'm told, support standard ANSI escape sequences, so you could choose to hard-code those sequences if Windows is the platform on which you want to run your version of this script. In many cases, you may prefer to use other existing Unix commands, such as lynx -dump -, to get richer formatting than this recipe provides. However, this recipe comes in quite handy when you find yourself on a system that has a Python installation but lacks such other helpful commands as lynx. See Also Library Reference and Python in a Nutshell docs on the formatter and htmllib modules; man tput on a Unix or Unix-like system for more information about the tput command. Chapter 2. Files Introduction Recipe 2.1. Reading from a File Recipe 2.2. Writing to a File Recipe 2.3. Searching and Replacing Text in a File Recipe 2.4. Reading a Specific Line from a File Recipe 2.5. Counting Lines in a File Recipe 2.6. Processing Every Word in a File Recipe 2.7. Using Random-Access Input/Output Recipe 2.8. Updating a Random-Access File Recipe 2.9. Reading Data from zip Files Recipe 2.10. Handling a zip File Inside a String Recipe 2.11. Archiving a Tree of Files into a Compressed tar File Recipe 2.12. Sending Binary Data to Standard Output Under Windows Recipe 2.13. Using a C++-like iostream Syntax Recipe 2.14. Rewinding an Input File to the Beginning Recipe 2.15. Adapting a File-like Object to a True File Object Recipe 2.16. Walking Directory Trees Recipe 2.17. Swapping One File Extension for Another Throughout a Directory Tree Recipe 2.18. Finding a File Given a Search Path Recipe 2.19. Finding Files Given a Search Path and a Pattern Recipe 2.20. Finding a File on the Python Search Path Recipe 2.21. Dynamically Changing the PythonSearch Path Recipe 2.22. Computing the Relative Path from One Directory to Another Recipe 2.23. Reading an Unbuffered Character in a Cross-Platform Way Recipe 2.24. Counting Pages of PDF Documents on Mac OS X Recipe 2.25. Changing File Attributes on Windows Recipe 2.26. Extracting Text from OpenOffice.org Documents Recipe 2.27. Extracting Text from Microsoft Word DocumentsRecipe 2.28. File Locking Using a Cross-Platform API Recipe 2.29. Versioning Filenames Recipe 2.30. Calculating CRC-64 Cyclic Redundancy Checks Introduction Credit: Mark Lutz, author of Programming Python and Python Quick Reference, co-author of Learning Python Behold the fileone of the first things that any reasonably pragmatic programmer reaches for in a programming language's toolbox. Because processing external files is a very real, tangible task, the quality of file-processing interfaces is a good way to assess the practicality of a programming tool. As the recipes in this chapter attest, Python shines in this task. Files in Python are supported in a variety of layers: from the built-in open function (a synonym for the standard file object type), to specialized tools in standard library modules such as os, to third-party utilities available on the Web. All told, Python's arsenal of file tools provides several powerful ways to access files in your scripts. File Basics In Python, a file object is an instance of built-in type file. The built-in function open creates and returns a file object. The first argument, a string, specifies the file's path (i.e., the filename preceded by an optional directory path). The second argument to open, also a string, specifies the mode in which to open the file. For example: input = open('data', 'r') output = open('/tmp/spam', 'w') open accepts a file path in which directories and files are separated by slash characters (/), regardless of the proclivities of the underlying operating system. On systems that don't use slashes, you can use a backslash character (\) instead, but there's no real reason to do so. Backslashes are harder to fit nicely in string literals, since you have to double them up or use "raw" strings. If the file path argument does not include the file's directory name, the file is assumed to reside in the current working directory (which is a disjoint concept from the Python module search path). For the mode argument, use 'r' to read the file in text mode; this is the default value and is commonly omitted, so that open is called with just one argument. Other common modes are 'rb' to read the file in binary mode, 'w' to create and write to the file in text mode, and 'wb' to create and write to the file in binary mode. A variant of 'r' that is sometimes precious is 'rU', which tells Python to read the file in text mode with "universal newlines": mode 'rU' can read text files independently of the line-termination convention the files are using, be it the Unix way, the Windows way, or even the (old) Mac way. (Mac OS X today is a Unix for all intents and purposes, but releases of Mac OS 9 and earlier, just a few years ago, were quite different.) The distinction between text mode and binary mode is important on non-Unix-like platforms because of the line-termination characters used on these systems. When you open a file in binary mode, Python knows that it doesn't need to worry about line-termination characters; it just moves bytes between the file and in-memory strings without any kind of translation. When you open a file in text mode on a non-Unix-like system, however, Python knows it must translate between the '\n' line-termination characters used in strings and whatever the current platform uses in the file itself. All of your Python code can always rely on '\n' as the line-termination character, as long as you properly indicate text or binary mode when you open the file. Once you have a file object, you perform all file I/O by calling methods of this object, as we'll discuss in a moment. When you're done with the file, you should finish by calling the closemethod on the object, to close the connection to the file: input.close( ) In short scripts, people often omit this step, as Python automatically closes the file when a file object is reclaimed during garbage collection (which in mainstream Python means the file is closed just about at once, although other important Python implementations, such as Jython and IronPython, have other, more relaxed garbage-collection strategies). Nevertheless, it is good programming practice to close your files as soon as possible, and it is especially a good idea in larger programs, which otherwise may be at more risk of having excessive numbers of uselessly open files lying about. Note that TRy/finally is particularly well suited to ensuring that a file gets closed, even when a function terminates due to an uncaught exception. To write to a file, use the write method: output.write(s) where s is a string. Think of s as a string of characters if output is open for text-mode writing, and as a string of bytes if output is open for binary-mode writing. Files have other writing- related methods, such as flush, to send any data being buffered, and writelines, to write a sequence of strings in a single call. However, write is by far the most commonly used method. Reading from a file is more common than writing to a file, and more issues are involved, so file objects have more reading methods than writing ones. The readline method reads and returns the next line from a text file. Consider the following loop: while True: line = input.readline( ) if not line: break process(line) This was once idiomatic Python but it is no longer the best way to read and process all of the lines from a file. Another dated alternative is to use the readlines method, which reads the whole file and returns a list of lines: for line in input.readlines( ): process(line) readlines is useful only for files that fit comfortably in physical memory. If the file is truly huge, readlines can fail or at least slow things down quite drastically (virtual memory fills up and the operating system has to start copying parts of physical memory to disk). In today's Python, just loop on the file object itself to get a line at a time with excellent memory and performance characteristics: for line in input: process(line) Of course, you don't always want to read a file line by line. You may instead want to read some or all of the bytes in the file, particularly if you've opened the file for binary-mode reading, where lines are unlikely to be an applicable concept. In this case, you can use the read method. When called without arguments, read reads and returns all the remaining bytes from the file. When read is called with an integer argument N, it reads and returns the next N bytes (or all the remaining bytes, if less than N bytes remain). Other methods worth mentioning are seek and tell, which support random access to files. These methods are normally used with binary files made up of fixed-length records.Portability and Flexibility On the surface, Python's file support is straightforward. However, before you peruse the code in this chapter, I want to underscore two aspects of Python's file support: code portability and interface flexibility. Keep in mind that most file interfaces in Python are fully portable across platform boundaries. It would be difficult to overstate the importance of this feature. A Python script that searches all files in a "directory" tree for a bit of text, for example, can be freely moved from platform to platform without source-code changes: just copy the script's source file to the new target machine. I do it all the timeso much so that I can happily stay out of operating system wars. With Python's portability, the underlying platform is almost irrelevant. Also, it has always struck me that Python's file-processing interfaces are not restricted to real, physical files. In fact, most file tools work with any kind of object that exposes the same interface as a real file object. Thus, a file reader cares only about read methods, and a file writer cares only about write methods. As long as the target object implements the expected protocol, all goes well. For example, suppose you have written a general file-processing function such as the following, meant to apply a passed-in function to each line of an input file: def scanner(fileobject, linehandler): for line in fileobject: linehandler(line) If you code this function in a module file and drop that file into a "directory" that's on your Python search path (sys.path), you can use it any time you need to scan a text file line by line, now or in the future. To illustrate, here is a client script that simply prints the first word of each line: from myutils import scanner def firstword(line): print line.split( )[0] file = open('data') scanner(file, firstword) So far, so good; we've just coded a small, reusable software component. But notice that there are no type declarations in the scanner function, only an interface constraintany object that is iterable line by line will do. For instance, suppose you later want to provide canned test input from a string object, instead of using a real, physical file. The standard StringIO module, and the equivalent but faster cStringIO, provide the appropriate wrapping and interface forgery: from cStringIO import StringIO from myutils import scanner def firstword(line): print line.split( )[0] string = StringIO('one\ntwo xxx\nthree\n') scanner(string, firstword) StringIO objects are plug-and-play compatible with file objects, so scanner takes its three lines of text from an in-memory string object, rather than a true external file. You don't need to change the scanner to make this workjust pass it the right kind of object. For more generality, you can even use a class to implement the expected interface instead: class MyStream(object): def _ _iter_ _(self): # grab and return text from wherever return iter(['a\n', 'b c d\n']) from myutils import scanner def firstword(line): print line.split( )[0] object = MyStream( ) scanner(object, firstword) This time, as scanner attempts to read the file, it really calls out to the _ _iter_ _ method you've coded in your class. In practice, such a method might use other Python standard tools to grab text from a variety of sources: an interactive user, a popup GUI input box, a shelve object, an SQL database, an XML or HTML page, a network socket, and so on. The point is that scanner doesn't know or care what type of object is implementing the interface it expects, or what that interface actually does. Object-oriented programmers know this deliberate naiveté as polymorphism. The type of the object being processed determines what an operation, such as the for-loop iteration in scanner, actually does. Everywhere in Python, object interfaces, rather than specific data types, are the unit of coupling. The practical effect is that functions are often applicable to a much broader range of problems than you might expect. This is especially true if you have a background in statically typed languages such as C or C++. It is almost as if we get C++ templates for free in Python. Code has an innate flexibility that is a by-product of Python's strong but dynamic typing. Of course, code portability and flexibility run rampant in Python development and are not really confined to file interfaces. Both are features of the language that are simply inherited by file- processing scripts. Other Python benefits, such as its easy scriptability and code readability, are also key assets when it comes time to change file-processing programs. But rather than extolling all of Python's virtues here, I'll simply defer to the wonderful recipes in this chapter and this book at large for more details. Enjoy! Recipe 2.1. Reading from a File Credit: Luther Blissett Problem You want to read text or data from a file. Solution Here's the most convenient way to read all of the file's contents at once into one long string: all_the_text = open('thefile.txt').read( ) # all text from a text file all_the_data = open('abinfile', 'rb').read( ) # all data from a binary file However, it is safer to bind the file object to a name, so that you can call close on it as soon as you're done, to avoid ending up with open files hanging around. For example, for a text file: file_object = open('thefile.txt') try: all_the_text = file_object.read( ) finally: file_object.close( ) You don't necessarily have to use the TRy/finally statement here, but it's a good idea to use it, because it ensures the file gets closed even when an error occurs during reading. The simplest, fastest, and most Pythonic way to read a text file's contents at once as a list of strings, one per line, is: list_of_all_the_lines = file_object.readlines( ) This leaves a '\n' at the end of each line; if you don't want that, you have alternatives, such as: list_of_all_the_lines = file_object.read( ).splitlines( ) list_of_all_the_lines = file_object.read( ).split('\n') list_of_all_the_lines = [L.rstrip('\n') for L in file_object] The simplest and fastest way to process a text file one line at a time is simply to loop on the file object with a for statement: for line in file_object: process line This approach also leaves a '\n' at the end of each line; you may remove it by starting the for loop's body with: line = line.rstrip('\n')or even, when you're OK with getting rid of trailing whitespace from each line (not just a trailing '\n'), the generally handier: line = line.rstrip( ) Discussion Unless the file you're reading is truly huge, slurping it all into memory in one gulp is often fastest and most convenient for any further processing. The built-in function open creates a Python file object (alternatively, you can equivalently call the built-in type file). You call the read method on that object to get all of the contents (whether text or binary) as a single long string. If the contents are text, you may choose to immediately split that string into a list of lines with the split method or the specialized splitlines method. Since splitting into lines is frequently needed, you may also call readlines directly on the file object for faster, more convenient operation. You can also loop directly on the file object, or pass it to callables that require an iterable, such as list or maxwhen thus treated as an iterable, a file object open for reading has the file's text lines as the iteration items (therefore, this should be done for text files only). This kind of line- by-line iteration is cheap in terms of memory consumption and fairly speedy too. On Unix and Unix-like systems, such as Linux, Mac OS X, and other BSD variants, there is no real distinction between text files and binary data files. On Windows and very old Macintosh systems, however, line terminators in text files are encoded, not with the standard '\n' separator, but with '\r\n' and '\r', respectively. Python translates these line-termination characters into '\n' on your behalf. This means that you need to tell Python when you open a binary file, so that it won't perform such translation. To do so, use 'rb' as the second argument to open. This is innocuous even on Unix-like platforms, and it's a good habit to distinguish binary files from text files even there, although it's not mandatory in that case. Such good habits will make your programs more immediately understandable, as well as more compatible with different platforms. If you're unsure about which line-termination convention a certain text file might be using, use 'rU' as the second argument to open, requesting universal endline translation. This lets you freely interchange text files among Windows, Unix (including Mac OS X), and old Macintosh systems, without worries: all kinds of line-ending conventions get mapped to '\n', whatever platform your code is running on. You can call methods such as read directly on the file object produced by the open function, as shown in the first snippet of the solution. When you do so, you no longer have a reference to the file object as soon as the reading operation finishes. In practice, Python notices the lack of a reference at once, and immediately closes the file. However, it is better to bind a name to the result of open, so that you can call close yourself explicitly when you are done with the file. This ensures that the file stays open for as short a time as possible, even on platforms such as Jython, IronPython, and other hypothetical future versions of Python, on which more advanced garbage-collection mechanisms might delay the automatic closing that the current version of C- based Python performs at once. To ensure that a file object is closed even if errors happen during its processing, the most solid and prudent approach is to use the try/finally statement: file_object = open('thefile.txt') try: for line in file_object: process line finally: file_object.close( )Be careful not to place the call to open inside the try clause of this try/finally statement (a rather common error among beginners). If an error occurs during the opening, there is nothing to close, and besides, nothing gets bound to name file_object, so you definitely don't want to call file_object.close()! If you choose to read the file a little at a time, rather than all at once, the idioms are different. Here's one way to read a binary file 100 bytes at a time, until you reach the end of the file: file_object = open('abinfile', 'rb') try: while True: chunk = file_object.read(100) if not chunk: break do_something_with(chunk) finally: file_object.close( ) Passing an argument N to the read method ensures that read will read only the next N bytes (or fewer, if the file is closer to the end). read returns the empty string when it reaches the end of the file. Complicated loops are best encapsulated as reusable generators. In this case, we can encapsulate the logic only partially, because a generator's yield keyword is not allowed in the try clause of a try/finally statement. Giving up on the assurance of file closing afforded by try/finally, we can therefore settle for: def read_file_by_chunks(filename, chunksize=100): file_object = open(filename, 'rb') while True: chunk = file_object.read(chunksize) if not chunk: break yield chunk file_object.close( ) Once this read_file_by_chunks generator is available, your application code to read and process a binary file by fixed-size chunks becomes extremely simple: for chunk in read_file_by_chunks('abinfile'): do_something_with(chunk) Reading a text file one line at a time is a frequent task. Just loop on the file object, as in: for line in open('thefile.txt', 'rU'): do_something_with(line) Here, too, in order to be 100% certain that no uselessly open file object will ever be left just hanging around, you may want to code this snippet in a more rigorously correct and prudent way: file_object = open('thefile.txt', 'rU'): try: for line in file_object: do_something_with(line) finally: file_object.close( )See Also Recipe 2.2; documentation for the open built-in function and file objects in the Library Reference and Python in a Nutshell. Recipe 2.2. Writing to a File Credit: Luther Blissett Problem You want to write text or data to a file. Solution Here is the most convenient way to write one long string to a file: open('thefile.txt', 'w').write(all_the_text) # text to a text file open('abinfile', 'wb').write(all_the_data) # data to a binary file However, it is safer to bind the file object to a name, so that you can call close on the file object as soon as you're done. For example, for a text file: file_object = open('thefile.txt', 'w') file_object.write(all_the_text) file_object.close( ) Often, the data you want to write is not in one big string, but in a list (or other sequence) of strings. In this case, you should use the writelines method (which, despite its name, is not limited to lines and works just as well with binary data as with text files!): file_object.writelines(list_of_text_strings) open('abinfile', 'wb').writelines(list_of_data_strings) Calling writelines is much faster than the alternatives of joining the strings into one big string (e.g., with ''.join) and then calling write, or calling write repeatedly in a loop. Discussion To create a file object for writing, you must always pass a second argument to open (or file)either 'w' to write textual data or 'wb' to write binary data. The same considerations detailed previously in Recipe 2.1 apply here, except that calling close explicitly is even more advisable when you're writing to a file rather than reading from it. Only by closing the file can you be reasonably sure that the data is actually on the disk and not still residing in some temporary buffer in memory. Writing a file a little at a time is even more common than reading a file a little at a time. You can just call write and/or writelines repeatedly, as each string or sequence of strings to write becomes ready. Each write operation appends data at the end of the file, after all the previously written data. When you're done, call the close method on the file object. If all the data is available at once, a single writelines call is faster and simpler. However, if the data becomes available a little at a time, it's better to call write as the data comes, than to build up a temporary list of pieces (e.g., with append) just in order to be able to write it all at once in theend with writelines. Reading and writing are quite different, with respect to the performance and convenience implications of operating "in bulk" versus operating a little at a time. When you open a file for writing with option 'w' (or 'wb'), any data that might already have been in the file is immediately destroyed; even if you close the file object immediately after opening it, you still end up with an empty file on the disk. If you want the data you're writing to be appended to the previous contents of the file, open the file with option 'a' (or 'ab') instead. More advanced options allow both reading and writing on the same open file objectin particular, see Recipe 2.8 for option 'r+b', which, in practice, is the only frequently used one out of all the advanced option strings. See Also Recipe 2.1; Recipe 2.8; documentation for the open built-in function and file objects in the Library Reference and Python in a Nutshell. Recipe 2.3. Searching and Replacing Text in a File Credit: Jeff Bauer, Adam Krieg Problem You need to change one string into another throughout a file. Solution String substitution is most simply performed by the replace method of string objects. The work here is to support reading from a specified file (or standard input) and writing to a specified file (or standard output): #!/usr/bin/env python import os, sys nargs = len(sys.argv) if not 3 <= nargs <= 5: print "usage: %s search_text replace_text [infile [outfile]]" % \ os.path.basename(sys.argv[0]) else: stext = sys.argv[1] rtext = sys.argv[2] input_file = sys.stdin output_file = sys.stdout if nargs > 3: 7 input_file = open(sys.argv[3]) if nargs > 4: output_file = open(sys.argv[4], 'w') for s in input_file: output_file.write(s.replace(stext, rtext)) output.close( ) input.close( ) Discussion This recipe is really simple, but that's what beautiful about itwhy do complicated stuff when simple stuff suffices? As indicated by the leading "shebang" line, the recipe is a simple main script, meaning a script meant to be run directly at a shell command prompt, as opposed to a module meant to be imported from elsewhere. The script looks at its arguments to determine the search text, the replacement text, the input file (defaulting to standard input), and the output file (defaulting to standard output). Then, it loops over each line of the input file, writing to the output file a copy of the line with the substitution performed on it. That's all! For accuracy, the script closes both files at the end. As long as an input file fits comfortably in memory in two copies (one before and one after the replacement, since strings are immutable), we could, with an increase in speed, operate on the entire input file's contents at once instead of looping. With today's low-end PCs typically containing at least 256 MB of memory, handling files of up to about 100 MB should not be a problem, and few text files are bigger than that. It suffices to replace the for loop with onesingle statement: output_file.write(input_file.read( ).replace(stext, rtext)) As you can see, that's even simpler than the loop used in the recipe. See Also Documentation for the open built-in function, file objects, and strings' replace method in the Library Reference and Python in a Nutshell. Recipe 2.4. Reading a Specific Line from a File Credit: Luther Blissett Problem You want to read from a text file a single line, given the line number. Solution The standard Python library linecache module makes this task a snap: import linecache theline = linecache.getline(thefilepath, desired_line_number) Discussion The standard linecache module is usually the optimal Python solution for this task. linecache is particularly useful when you have to perform this task repeatedly for several lines in a file, since linecache caches information to avoid uselessly repeating work. When you know that you won't be needing any more lines from the cache for a while, call the module's clearcache function to free the memory used for the cache. You can also use checkcache if the file may have changed on disk and you must make sure you are getting the updated version. linecache reads and caches all of the text file whose name you pass to it, so, if it's a very large file and you need only one of its lines, linecache may be doing more work than is strictly necessary. Should this happen to be a bottleneck for your program, you may get an increase in speed by coding an explicit loop, encapsulated within a function, such as: def getline(thefilepath, desired_line_number): if desired_line_number < 1: return '' for current_line_number, line in enumerate(open(thefilepath, 'rU')): if current_line_number == desired_line_number-1: return line return '' The only detail requiring attention is that enumerate counts from 0, so, since we assume the desired_line_number argument counts from 1, we need the -1 in the == comparison. See Also Documentation for the linecache module in the Library Reference and Python in a Nutshell; Perl Cookbook recipe 8.8. Recipe 2.5. Counting Lines in a File Credit: Luther Blissett Problem You need to compute the number of lines in a file. Solution The simplest approach for reasonably sized files is to read the file as a list of lines, so that the count of lines is the length of the list. If the file's path is in a string bound to a variable named thefilepath, all the code you need to implement this approach is: count = len(open(thefilepath, 'rU').readlines( )) For a truly huge file, however, this simple approach may be very slow or even fail to work. If you have to worry about humongous files, a loop on the file always works: count = -1 for count, line in enumerate(open(thefilepath, 'rU')): pass count += 1 A tricky alternative, potentially faster for truly humongous files, for when the line terminator is '\n' (or has '\n' as a substring, as happens on Windows): count = 0 thefile = open(thefilepath, 'rb') while True: buffer = thefile.read(8192*1024) if not buffer: break count += buffer.count('\n') thefile.close( ) The 'rb' argument to open is necessary if you're after speedwithout that argument, this snippet might be very slow on Windows. Discussion When an external program counts a file's lines, such as wc -l on Unix-like platforms, you can of course choose to use that (e.g., via os.popen). However, it's generally simpler, faster, and more portable to do the line-counting in your own program. You can rely on almost all text files having a reasonable size, so that reading the whole file into memory at once is feasible. For all such normal files, the len of the result of readlines gives you the count of lines in the simplest way. If the file is larger than available memory (say, a few hundred megabytes on a typical PC today),the simplest solution can become unacceptably slow, as the operating system struggles to fit the file's contents into virtual memory. It may even fail, when swap space is exhausted and virtual memory can't help any more. On a typical PC, with 256MB RAM and virtually unlimited disk space, you should still expect serious problems when you try to read into memory files above, say, 1 or 2 GB, depending on your operating system. (Some operating systems are much more fragile than others in handling virtual-memory issues under such overly stressed load conditions.) In this case, looping on the file object, as shown in this recipe's Solution, is better. The enumerate built-in keeps the line count without your code having to do it explicitly. Counting line-termination characters while reading the file by bytes in reasonably sized chunks is the key idea in the third approach. It's probably the least immediately intuitive, and it's not perfectly cross-platform, but you might hope that it's fastest (e.g., when compared with recipe 8.2 in the Perl Cookbook). However, in most cases, performance doesn't really matter all that much. When it does matter, the time-sink part of your program might not be what your intuition tells you it is, so you should never trust your intuition in this matterinstead, always benchmark and measure. For example, consider a typical Unix syslog file of middling size, a bit over 18 MB of text in 230,000 lines: [situ@tioni nuc]$ wc nuc 231581 2312730 18508908 nuc And consider the following testing-and-benchmark framework script, bench.py: import time def timeo(fun, n=10): start = time.clock( ) for i in xrange(n): fun( ) stend = time.clock( ) thetime = stend-start return fun._ _name_ _, thetime import os def linecount_w( ): return int(os.popen('wc -l nuc').read( ).split( )[0]) def linecount_1( ): return len(open('nuc').readlines( )) def linecount_2( ): count = -1 for count, line in enumerate(open('nuc')): pass return count+1 def linecount_3( ): count = 0 thefile = open('nuc', 'rb') while True: buffer = thefile.read(65536) if not buffer: break count += buffer.count('\n') return count for f in linecount_w, linecount_1, linecount_2, linecount_3: print f._ _name_ _, f( ) for f in linecount_1, linecount_2, linecount_3: print "%s: %.2f"%timeo(f) First, I print the line-counts obtained by all methods, thus ensuring that no anomaly or error has occurred (counting tasks are notoriously prone to off-by-one errors). Then, I run each alternative 10 times, under the control of the timing function timeo, and look at the results. Here they are, on the old but reliable machine I measured them on: [situ@tioni nuc]$ python -O bench.pylinecount_w 231581 linecount_1 231581 linecount_2 231581 linecount_3 231581 linecount_1: 4.84 linecount_2: 4.54 linecount_3: 5.02 As you can see, the performance differences hardly matter: your users will never even notice a difference of 10% or so in one auxiliary task. However, the fastest approach (for my particular circumstances, on an old but reliable PC running a popular Linux distribution, and for this specific benchmark) is the humble loop-on-every-line technique, while the slowest one is the fancy, ambitious technique that counts line terminators by chunks. In practice, unless I had to worry about files of many hundreds of megabytes, I'd always use the simplest approach (i.e., the first one presented in this recipe). Measuring the exact performance of code snippets (rather than blindly using complicated approaches in the hope that they'll be faster) is very importantso important, indeed, that the Python Standard Library includes a module, timeit, specifically designed for such measurement tasks. I suggest you use timeit, rather than coding your own little benchmarks as I have done here. The benchmark I just showed you is one I've had around for years, since well before timeit appeared in the standard Python library, so I think I can be forgiven for not using timeit in this specific case! See Also The Library Reference and Python in a Nutshell sections on file objects, the enumerate built-in, os.popen, and the time and timeit modules; Perl Cookbook recipe 8.2. Recipe 2.6. Processing Every Word in a File Credit: Luther Blissett Problem You need to do something with each and every word in a file. Solution This task is best handled by two nested loops, one on lines and another on the words in each line: for line in open(thefilepath): for word in line.split( ): dosomethingwith(word) The nested for statement's header implicitly defines words as sequences of nonspaces separated by sequences of spaces (just as the Unix program wc does). For other definitions of words, you can use regular expressions. For example: import re re_word = re.compile(r"[\w'-]+") for line in open(thefilepath): for word in re_word.finditer(line): dosomethingwith(word.group(0)) In this case, a word is defined as a maximal sequence of alphanumerics, hyphens, and apostrophes. Discussion If you want to use other definitions of words, you will obviously need different regular expressions. The outer loop, on all lines in the file, won't change. It's often a good idea to wrap iterations as iterator objects, and this kind of wrapping is most commonly and conveniently obtained by coding simple generators: def words_of_file(thefilepath, line_to_words=str.split): the_file = open(thefilepath): for line in the_file: for word in line_to_words(line): yield word the_file.close( ) for word in words_of_file(thefilepath): dosomethingwith(word) This approach lets you separate, cleanly and effectively, two different concerns: how to iterateover all items (in this case, words in a file) and what to do with each item in the iteration. Once you have cleanly encapsulated iteration concerns in an iterator object (often, as here, a generator), most of your uses of iteration become simple for statements. You can often reuse the iterator in many spots in your program, and if maintenance is ever needed, you can perform that maintenance in just one placethe definition of the iteratorrather than having to hunt for all uses. The advantages are thus very similar to those you obtain in any programming language by appropriately defining and using functions, rather than copying and pasting pieces of code all over the place. With Python's iterators, you can get these reuse advantages for all of your looping-control structures, too. We've taken the opportunity afforded by the refactoring of the loop into a generator to perform two minor enhancementsensuring the file is explicitly closed, which is always a good idea, and generalizing the way each line is split into words (defaulting to the split method of string objects, but leaving a door open to more generality). For example, when we need words as defined by a regular expression, we can code another wrapper on top of words_of_file thanks to this "hook": import re def words_by_re(thefilepath, repattern=r"[\w'-]+"): wre = re.compile(repattern) def line_to_words(line): for mo in wre.finditer(line): return mo.group(0) return words_of_file(thefilepath, line_to_words) Here, too, we supply a reasonable default for the regular expression pattern defining a word but still make it easy to pass a different value in those cases in which different definitions are necessary. Excessive generalization is a pernicious temptation, but a little tasteful generalization suggested by experience will most often amply repay the modest effort it requires. Having a function accept an optional argument, while providing the most likely value for the argument as the default value, is among the simplest and handiest ways to implement this modest and often worthwhile kind of generalization. See Also Chapter 19 for more on iterators and generators; Library Reference and Python in a Nutshell on file objects and the re module; Perl Cookbook recipe 8.3. Recipe 2.7. Using Random-Access Input/Output Credit: Luther Blissett Problem You want to read a binary record from somewhere inside a large file of fixed-length records, without reading a record at a time to get there. Solution The byte offset of the start of a record in the file is the size of a record, in bytes, multiplied by the progressive number of the record (counting from 0). So, you can just seek right to the proper spot, then read the data. For example, to read the seventh record from a binary file where each record is 48 bytes long: thefile = open('somebinfile', 'rb') record_size = 48 record_number = 6 thefile.seek(record_size * record_number) buffer = thefile.read(record_size) Note that the record_number of the seventh record is 6: record numbers count from zero! Discussion This approach works only on files (generally binary ones) defined in terms of records that are all the same fixed size in bytes; it doesn't work on normal text files. For clarity, the recipe shows the file being opened for reading as a binary file by passing 'rb' as the second argument to open, just before the seek. As long as the file object is open for reading as a binary file, you can perform as many seek and read operations as you need, before eventually closing the file againyou don't necessarily open the file just before performing a seek on it. See Also The section of the Library Reference and Python in a Nutshell on file objects; Perl Cookbook recipe 8.12. Recipe 2.8. Updating a Random-Access File Credit: Luther Blissett Problem You want to read a binary record from somewhere inside a large file of fixed-length records, change some or all of the values of the record's fields, and write the record back. Solution Read the record, unpack it, perform whatever computations you need for the update, pack the fields back into the record, seek to the start of the record again, write it back. Phew. Faster to code than to say: import struct format_string = '8l' # e.g., say a record is 8 4-byte integers thefile = open('somebinfile', 'r+b') record_size = struct.calcsize(format_string) thefile.seek(record_size * record_number) buffer = thefile.read(record_size) fields = list(struct.unpack(format_string, buffer)) # Perform computations, suitably modifying fields, then: buffer = struct.pack(format_string, *fields) thefile.seek(record_size * record_number) thefile.write(buffer) thefile.close( ) Discussion This approach works only on files (generally binary ones) defined in terms of records that are all the same, fixed size; it doesn't work on normal text files. Furthermore, the size of each record must be that defined by a struct format string, as shown in the recipe's code. A typical format string, for example, might be '8l', to specify that each record is made up of eight four-byte integers, each to be interpreted as a signed value and unpacked into a Python int. In this case, the fields variable in the recipe would be bound to a list of eight ints. Note that struct.unpack returns a tuple. Because tuples are immutable, the computation would have to rebind the entire fields variable. A list is mutable, so each field can be rebound as needed. Thus, for convenience, we explicitly ask for a list when we bind fields. Make sure, however, not to alter the length of the list. In this case, it needs to remain composed of exactly eight integers, or the struct.pack call will raise an exception when we call it with a format_string of '8l'. Also, this recipe is not suitable when working with records that are not all of the same, unchanging length. To seek back to the start of the record, instead of using the record_size*record_number offset again, you may choose to do a relative seek: thefile.seek(-record_size, 1) The second argument to the seek method (1) tells the file object to seek relative to the currentposition (here, so many bytes back, because we used a negative number as the first argument). seek's default is to seek to an absolute offset within the file (i.e., from the start of the file). You can also explicitly request this default behavior by calling seek with a second argument of 0. You don't need to open the file just before you do the first seek, nor do you need to close it right after the write. Once you have a file object that is correctly opened (i.e., for updating and as a binary rather than a text file), you can perform as many updates on the file as you want before closing the file again. These calls are shown here to emphasize the proper technique for opening a file for random-access updates and the importance of closing a file when you are done with it. The file needs to be opened for updating (i.e., to allow both reading and writing). That's what the 'r+b' argument to open means: open for reading and writing, but do not implicitly perform any transformations on the file's contents because the file is a binary one. (The 'b' part is unnecessary but still recommended for clarity on Unix and Unix-like systems. However, it's absolutely crucial on other platforms, such as Windows.) If you're creating the binary file from scratch, but you still want to be able to go back, reread, and update some records without closing and reopening the file, you can use a second argument of 'w+b' instead. However, I have never witnessed this strange combination of requirements; binary files are normally first created (by opening them with 'wb', writing data, and closing the file) and later reopened for updating with 'r+b'. While this approach is normally useful only on a file whose records are all the same size, another, more advanced possibility exists: a separate "index file" that provides the offset and length of each record inside the "data file". Such indexed sequential access approaches aren't much in fashion any more, but they used to be very important. Nowadays, one meets just about only text files (of many kinds, more and more often XML ones), databases, and occasional binary files with fixed-length records. Still, if you do need to access an indexed sequential binary file, the code is quite similar to that shown in this recipe, except that you must obtain the record_size and the offset argument to pass to thefile.seek by reading them from the index file, rather than computing them yourself as shown in this recipe's Solution. See Also The sections of the Library Reference and Python in a Nutshell on file objects and the struct module; Perl Cookbook recipe 8.13. Recipe 2.9. Reading Data from zip Files Credit: Paul Prescod, Alex Martelli Problem You want to directly examine some or all of the files contained in an archive in zip format, without expanding them on disk. Solution zip files are a popular, cross-platform way of archiving files. The Python Standard Library comes with a zipfile module to access such files easily: import zipfile z = zipfile.ZipFile("zipfile.zip", "r") for filename in z.namelist( ): print 'File:', filename, bytes = z.read(filename) print 'has', len(bytes), 'bytes' Discussion Python can work directly with data in zip files. You can look at the list of items in the archive's directory and work with the "data file"s themselves. This recipe is a snippet that lists all of the names and content lengths of the files included in the zip archive zipfile.zip. The zipfile module does not currently handle multidisk zip files nor zip files with appended comments. Take care to use r as the flag argument, not rb, which might seem more natural (e.g., on Windows). With ZipFile, the flag is not used the same way when opening a file, and rb is not recognized. The r flag handles the inherently binary nature of all zip files on all platforms. When a zip file contains some Python modules (meaning .py or preferably .pyc files), possibly in addition to other (data) files, you can add the file's path to Python's sys.path and then use the import statement to import modules from the zip file. Here's a toy, self-contained, purely demonstrative example that creates such a zip file on the fly, imports a module from it, then removes itall just to show you how it's done: import zipfile, tempfile, os, sys handle, filename = tempfile.mkstemp('.zip') os.close(handle) z = zipfile.ZipFile(filename, 'w') z.writestr('hello.py', 'def f( ): return "hello world from "+_ _file_ _\n') z.close( ) sys.path.insert(0, filename) import hello print hello.f( ) os.unlink(filename)Running this script emits something like: hello world from /tmp/tmpESVzeY.zip/hello.py Besides illustrating Python's ability to import from a zip file, this snippet also shows how to make (and later remove) a temporary file, and how to use the writestr method to add a member to a zip file without placing that member into a disk file first. Note that the path to the zip file from which you import is treated somewhat like a directory. (In this specific example run, that path is /tmp/tmpESVzeY.zip, but of course, since we're dealing with a temporary file, the exact value of the path can change at each run, depending also on your platform.) In particular, the _ _file_ _ global variable, within the module hello, which is imported from the zip file, has a value of /tmp/tmpESVzeY.zip/hello.pya pseudo-path, made up of the zip file's path seen as a "directory" followed by the relative path of hello.py within the zip file. If you import from a zip file a module that computes paths relative to itself in order to get to data files, you need to adapt the module to this effect, because you cannot just open such a "pseudo-path" to get a file object: rather, to read or write files inside a zip file, you must use functions from standard library module zipfile, as shown in the solution. For more information about importing modules from a zip file, see Recipe 16.12. While that recipe is Unix-specific, the information in the recipe's Discussion about importing from zip files is also valid for Windows. See Also Documentation for the zipfile module in the Library Reference and Python in a Nutshell; modules tempfile, os, sys; for archiving a tree of files, see Recipe 2.11; for more information about importing modules from a zip file, Recipe 16.12. Recipe 2.10. Handling a zip File Inside a String Credit: Indyana Jones Problem Your program receives a zip file as a string of bytes in memory, and you need to read the information in this zip file. Solution Solving this kind of problem is exactly what standard library module cStringIO is for: import cStringIO, zipfile class ZipString(ZipFile): def _ _init_ _(self, datastring): ZipFile._ _init_ _(self, cStringIO.StringIO(datastring)) Discussion I often find myself faced with this taskfor example, zip files coming from BLOB fields in a database or ones received from a network connection. I used to save such binary data to a temporary file, then open the file with the standard library module zipfile. Of course, I had to ensure I deleted the temporary file when I was done. Then I thought of using the standard library module cStringIO for the purpose . . . and never looked back. Module cStringIO lets you wrap a string of bytes so it can be accessed as a file object. You can also do things the other way around, writing into a cStringIO.StringIO instance as if it were a file object, and eventually recovering its contents as a string of bytes. Most Python modules that take file objects don't check whether you're passing an actual filerather, any file-like object will do; the module's code just calls on the object whatever file methods it needs. As long as the object supplies those methods and responds correctly when they're called, everything just works. This demonstrates the awesome power of signature-based polymorphism and hopefully teaches why you should almost never type-test (utter such horrors as if type(x) is y, or even just the lesser horror if isinstance(x, y)) in your own code! A few low-level modules, such as marshal, are unfortunately adamant about using "true" files, but zipfile isn't, and this recipe shows how simple it makes your life! If you are using a version of Python that is different from the mainstream C-coded one, known as "CPython", you may not find module cStringIO in the standard library. The leading c in the name of the module indicates that it's a C-specific module, optimized for speed but not guaranteed to be in the standard library for other compliant Python implementations. Several such alternative implementations include both production-quality ones (such as Jython, which is coded in Java and runs on a JVM) and experimental ones (such as pypy, which is coded in Python and generates machine code, and IronPython, which is coded in C# and runs on Microsoft's .NET CLR). Not to worry: the Python Standard Library always includes module StringIO, which is coded in pure Python (and thus is usable from any compliant implementation of Python), and implements the same functionality as module cStringIO (albeit not quite as fast, at least on the mainstream CPython implementation). You just need to alter your import statement a bit to make sure you get cStringIO when available and StringIO otherwise. For example, this recipe might become:import zipfile try: from cStringIO import StringIO except ImportError: from StringIO import StringIO class ZipString(ZipFile): def _ _init_ _(self, datastring): ZipFile._ _init_ _(self, StringIO(datastring)) With this modification, the recipe becomes useful in Jython, and other, alternative implementations. See Also Modules zipfile and cStringIO in the Library Reference and Python in a Nutshell; Jython is at http://www.jython.org/; pypy is at http://codespeak.net/pypy/; IronPython is at http://ironpython.com/. Recipe 2.11. Archiving a Tree of Files into a Compressed tar File Credit: Ed Gordon, Ravi Teja Bhupatiraju Problem You need to archive all of the files and folders in a subtree into a tar archive file, compressing the data with either the popular gzip approach or the higher-compressing bzip2 approach. Solution The Python Standard Library's tarfile module directly supports either kind of compression: you just need to specify the kind of compression you require, as part of the option string that you pass when you call tarfile.TarFile.open to create the archive file. For example: import tarfile, os def make_tar(folder_to_backup, dest_folder, compression='bz2'): if compression: dest_ext = '.' + compression else: dest_ext = '' arcname = os.path.basename(folder_to_backup) dest_name = '%s.tar%s' % (arcname, dest_ext) dest_path = os.path.join(dest_folder, dest_name) if compression: dest_cmp = ':' + compression else: dest_cmp = '' out = tarfile.TarFile.open(dest_path, 'w'+dest_cmp) out.add(folder_to_backup, arcname) out.close( ) return dest_path Discussion You can pass, as argument compression to function make_tar, the string 'gz' to get gzip compression instead of the default bzip2, or you can pass the empty string '' to get no compression at all. Besides making the file extension of the result either .tar, .tar.gz, or .tar.bz2, as appropriate, your choice for the compression argument determines which string is passed as the second argument to tarfile.TarFile.open: 'w', when you want no compression, or 'w:gz' or 'w:bz2' to get two kinds of compression. Class tarfile.TarFile offers several other classmethods, besides open, which you could use to generate a suitable instance. I find open handier and more flexible because it takes the compression information as part of the mode string argument. However, if you want to ensure bzip2 compression is used unconditionally, for example, you could choose to call classmethod bz2open instead. Once we have an instance of class tarfile.TarFile that is set to use the kind of compressionwe desire, the instance's method add does all we require. In particular, when string folder_to_backup names a "directory" (or folder), rather than an ordinary file, add recursively adds all of the subtree rooted in that directory. If on some other occasion, we wanted to change this behavior to get precise control on what is archived, we could pass to add an additional named argument recursive=False to switch off this implicit recursion. After calling add, all that's left for function make_tar to do is to close the TarFile instance and return the path on which the tar file has been written, just in case the caller needs this information. See Also Library Reference docs on module tarfile. Recipe 2.12. Sending Binary Data to Standard Output Under Windows Credit: Hamish Lawson Problem You want to send binary data (e.g., an image) to stdout under Windows. Solution That's what the setmode function, in the platform-dependent (Windows-only) msvcrt module in the Python Standard Library, is for: import sys if sys.platform == "win32": import os, msvcrt msvcrt.setmode(sys.stdout.fileno( ), os.O_BINARY) You can now call sys.stdout.write with any bytestring as the argument, and the bytestring will go unmodified to standard output. Discussion While Unix doesn't make (or need) a distinction between text and binary modes, if you are reading or writing binary data, such as an image, under Windows, the file must be opened in binary mode. This is a problem for programs that write binary data to standard output (as a CGI script, for example, could be expected to do), because Python opens the sys.stdout file object on your behalf, normally in text mode. You can have stdout opened in binary mode instead by supplying the -u command-line option to the Python interpreter. For example, if you know your CGI script will be running under the Apache web server, as the first line of your script, you can use something like: #! c:/python23/python.exe -u assuming you're running under Python 2.3 with a standard installation. Unfortunately, you may not always be able to control the command line under which your script will be started. The approach taken in this recipe's "Solution" offers a workable alternative. The setmode function provided by the Windows-specific msvcrt module lets you change the mode of stdout's underlying file descriptor. By using this function, you can ensure from within your program that sys.stdout gets set to binary mode. See Also Documentation for the msvcrt module in the Library Reference and Python in a Nutshell. Recipe 2.13. Using a C++-like iostream Syntax Credit: Erik Max Francis Problem You like the C++ approach to I/O, based on ostreams and manipulators (special objects that cause special effects on a stream when inserted in it) and want to use it in your Python programs. Solution Python lets you overload operators by having your classes define special methods (i.e., methods whose names start and end with two underscores). To use << for output, as you do in C++, you just need to code an output stream class that defines the special method _ _lshift_ _: class IOManipulator(object): def _ _init_ _(self, function=None): self.function = function def do(self, output): self.function(output) def do_endl(stream): stream.output.write('\n') stream.output.flush( ) endl = IOManipulator(do_endl) class OStream(object): def _ _init_ _(self, output=None): if output is None: import sys output = sys.stdout self.output = output self.format = '%s' def _ _lshift_ _(self, thing): ''' the special method which Python calls when you use the << operator and the left-hand operand is an OStream ''' if isinstance(thing, IOManipulator): thing.do(self) else: self.output.write(self.format % thing) self.format = '%s' return self def example_main( ): cout = OStream( ) cout<< "The average of " << 1 << " and " << 3 << " is " << (1+3)/2 <>somewhere, "The average of %d and %d is %f\n" % (1, 3, (1+3)/2) which is the "Python-native" alternative (looking a lot like C in this case). It depends in part on whether you're more used to C++ or to C. In any case, this recipe gives you a choice! Even if you don't end up using this particular approach, it's still interesting to see how simple operator overloading is in Python. See Also Library Reference and Python in a Nutshell docs on file objects and special methods such as _ _lshift_ _; Recipe 4.20 implements a Python version of C's printf function. Recipe 2.14. Rewinding an Input File to the Beginning Credit: Andrew Dalke Problem You need to make an input file object (with data coming from a socket or other input file handle) rewindable back to the beginning so you can read it over. Solution Wrap the file object into a suitable class: from cStringIO import StringIO class RewindableFile(object): """ Wrap a file handle to allow seeks back to the beginning. """ def _ _init_ _(self, input_file): """ Wraps input_file into a file-like object with rewind. """ self.file = input_file self.buffer_file = StringIO( ) self.at_start = True try: self.start = input_file.tell( ) except (IOError, AttributeError): self.start = 0 self._use_buffer = True def seek(self, offset, whence=0): """ Seek to a given byte position. Must be: whence == 0 and offset == self.start """ if whence != 0: raise ValueError("whence=%r; expecting 0" % (whence,)) if offset != self.start: raise ValueError("offset=%r; expecting %s" % (offset, self.start)) self.rewind( ) def rewind(self): """ Simplified way to seek back to the beginning. """ self.buffer_file.seek(0) self.at_start = True def tell(self): """ Return the current position of the file (must be at start). """ if not self.at_start: raise TypeError("RewindableFile can't tell except at start of file") return self.start def _read(self, size): if size < 0: # read all the way to the end of the file y = self.file.read( ) if self._use_buffer: self.buffer_file.write(y) return self.buffer_file.read( ) + y elif size == 0: # no need to actually read the empty string return "" x = self.buffer_file.read(size) if len(x) < size: y = self.file.read(size - len(x)) if self._use_buffer: self.buffer_file.write(y) return x + y return x def read(self, size=-1): """ Read up to 'size' bytes from the file. Default is -1, which means to read to end of file. """ x = self._read(size) if self.at_start and x: self.at_start = False self._check_no_buffer( ) return x def readline(self): """ Read a line from the file. """ # Can we get it out of the buffer_file? s = self.buffer_file.readline( ) if s[-1:] == "\n": return s # No, so read a line from the input file t = self.file.readline( ) if self._use_buffer: self.buffer_file.write(t) self._check_no_buffer( ) return s + t def readlines(self): """read all remaining lines from the file""" return self.read( ).splitlines(True) def _check_no_buffer(self): # If 'nobuffer' has been called and we're finished with the buffer file, # get rid of the buffer, redirect everything to the original input file. if not self._use_buffer and \ self.buffer_file.tell( ) == len(self.buffer_file.getvalue( )): # for top performance, we rebind all relevant methods in self for n in 'seek tell read readline readlines'.split( ): setattr(self, n, getattr(self.file, n, None)) del self.buffer_file def nobuffer(self): """tell RewindableFile to stop using the buffer once it's exhausted""" self._use_buffer = False Discussion Sometimes, data coming from a socket or other input file handle isn't what it was supposed to be. For example, suppose you are reading from a buggy server, which is supposed to return an XML stream, but sometimes returns an unformatted error message instead. (This scenario often occurs because many servers don't handle incorrect input very well.) This recipe's RewindableFile class helps you solve this problem. r = RewindableFile(f) wraps the original input stream f into a "rewindable file" instance r which essentially mimics f's behavior but also provides a buffer. Read requests to r are forwarded to f, and the data thus read gets appended to a buffer, then returned to the caller. The buffer contains all the data read so far. r can be told to rewind, meaning to seek back to the start position. The next read request willcome from the buffer, until the buffer has been read, in which case it gets the data from the input stream again. The newly read data is also appended to the buffer. When buffering is no longer needed, call the nobuffer method of r. This tells r that, once it's done reading the buffer's current contents, it can throw the buffer away. After nobuffer is called, the behavior of seek is no longer defined. For example, suppose you have a server that gives either an error message of the form ERROR: cannot do that, or an XML data stream, starting with '' % sys.argv[0] sys.exit(1) matches = list(all_files(sys.argv[1], os.environ['PATH'])) print '%d match:' % len(matches) for match in matches: print match See Also Recipe 2.18 for a simpler approach to find the first file with a specified name along the path; Library Reference and Python in a Nutshell docs for modules os and glob. Recipe 2.20. Finding a File on the Python Search Path Credit: Mitch Chapman Problem A large Python application includes resource files (e.g., Glade project files, SQL templates, and images) as well as Python packages. You want to store these associated files together with the Python packages that use them. Solution You need to be able to look for either files or directories along Python's sys.path: import sys, os class Error(Exception): pass def _find(pathname, matchFunc=os.path.isfile): for dirname in sys.path: candidate = os.path.join(dirname, pathname) if matchFunc(candidate): return candidate raise Error("Can't find file %s" % pathname) def findFile(pathname): return _find(pathname) def findDir(path): return _find(path, matchFunc=os.path.isdir) Discussion Larger Python applications consist of sets of Python packages and associated sets of resource files. It's convenient to store these associated files together with the Python packages that use them, and it's easy to do so if you use this variation on the previous Recipe 2.18 to find files or directories with pathnames relative to the Python search path. See Also Recipe 2.18; documentation for the os module in the Library Reference and Python in a Nutshell. Recipe 2.21. Dynamically Changing the PythonSearch Path Credit: Robin Parmar Problem Modules must be on the Python search path before they can be imported, but you don't want to set a huge permanent path because that slows performanceso, you want to change the path dynamically. Solution We simply conditionally add a "directory" to Python's sys.path, carefully checking to avoid duplication: def AddSysPath(new_path): """ AddSysPath(new_path): adds a "directory" to Python's sys.path Does not add the directory if it does not exist or if it's already on sys.path. Returns 1 if OK, -1 if new_path does not exist, 0 if it was already on sys.path. """ import sys, os # Avoid adding nonexistent paths if not os.path.exists(new_path): return -1 # Standardize the path. Windows is case-insensitive, so lowercase # for definiteness if we are on Windows. new_path = os.path.abspath(new_path) if sys.platform == 'win32': new_path = new_path.lower( ) # Check against all currently available paths for x in sys.path: x = os.path.abspath(x) if sys.platform == 'win32': x = x.lower( ) if new_path in (x, x + os.sep): return 0 sys.path.append(new_path) # if you want the new_path to take precedence over existing # directories already in sys.path, instead of appending, use: # sys.path.insert(0, new_path) return 1 if _ _name_ _ == '_ _main_ _': # Test and show usage import sys print 'Before:' for x in sys.path: print x if sys.platform == 'win32': print AddSysPath('c:\\Temp') print AddSysPath('c:\\temp') else: print AddSysPath('/usr/lib/my_modules') print 'After:' for x in sys.path: print x Discussion Modules must be in directories that are on the Python search path before they can be imported, but we don't want to have a huge permanent path because doing so slows down every import performed by every Python script and application. This simple recipe dynamically adds a "directory" to the path, but only if that directory exists and was not already on sys.path. sys.path is a list, so it's easy to add directories to its end, using sys.path.append. Every import performed after such an append will automatically look in the newly added directory if it cannot be satisfied from earlier ones. As indicated in the Solution, you can alternatively use sys.path.insert(0, . . . so that the newly added directory is searched before ones that were already in sys.path. It's no big deal if sys.path ends up with some duplicates or if a nonexistent directory is accidentally appended to it; Python's import statement is clever enough to shield itself against such issues. However, each time such a problem occurs at import time (e.g., from duplicate unsuccessful searches, errors from the operating system that need to be handled gracefully, etc.), a small price is paid in terms of performance. To avoid uselessly paying such a price, this recipe does a conditional addition to sys.path, never appending any directory that doesn't exist or is already in sys.path. Directories appended by this recipe stay in sys.path only for the duration of this program's run, just like any other dynamic alteration you might do to sys.path. See Also Documentation for the sys and os.path modules in the Library Reference and Python in a Nutshell. Recipe 2.22. Computing the Relative Path from One Directory to Another Credit: Cimarron Taylor, Alan Ezust Problem You need to know the relative path from one directory to anotherfor example, to create a symbolic link or a relative reference in a URL. Solution The simplest approach is to split paths into lists of directories, then work on the lists. Using a couple of auxiliary and somewhat generic helper functions, we could code: import os, itertools def all_equal(elements): ''' return True if all the elements are equal, otherwise False. ''' first_element = elements[0] for other_element in elements[1:]: if other_element != first_element: return False return True def common_prefix(*sequences): ''' return a list of common elements at the start of all sequences, then a list of lists that are the unique tails of each sequence. ''' # if there are no sequences at all, we're done if not sequences: return [ ], [ ] # loop in parallel on the sequences common = [ ] for elements in itertools.izip(*sequences): # unless all elements are equal, bail out of the loop if not all_equal(elements): break # got one more common element, append it and keep looping common.append(elements[0]) # return the common prefix and unique tails return common, [ sequence[len(common):] for sequence in sequences ] def relpath(p1, p2, sep=os.path.sep, pardir=os.path.pardir): ''' return a relative path from p1 equivalent to path p2. In particular: the empty string, if p1 == p2; p2, if p1 and p2 have no common prefix. ''' common, (u1, u2) = common_prefix(p1.split(sep), p2.split(sep)) if not common: return p2 # leave path absolute if nothing at all in common return sep.join( [pardir]*len(u1) + u2 ) def test(p1, p2, sep=os.path.sep): ''' call function relpath and display arguments and results. ''' print "from", p1, "to", p2, " -> ", relpath(p1, p2, sep) if _ _name_ _ == '_ _main_ _': test('/a/b/c/d', '/a/b/c1/d1', '/') test('/a/b/c/d', '/a/b/c/d', '/') test('c:/x/y/z', 'd:/x/y/z', '/') Discussion The workhorse in this recipe is the simple but very general function common_prefix, which, given any N sequences, returns their common prefix and a list of their respective unique tails. To compute the relative path between two given paths, we can ignore their common prefix. We need only the appropriate number of move-up markers (normally, os.path.pardire.g., ../ on Unix-like systems; we need as many of them as the length of the unique tail of the starting path) followed by the unique tail of the destination path. So, function relpath splits the paths into lists of directories, calls common_prefix, and then performs exactly the construction just described. common_prefix centers on the loop for elements in itertools.izip(*sequences), relying on the fact that izip ends with the shortest of the iterables it's zipping. The body of the loop only needs to prematurely terminate the loop as soon as it meets a tuple of elements (coming one from each sequence, per izip's specifications) that aren't all equal, and to keep track of the elements that are equal by appending one of them to list common. Once the loop is done, all that's left to prepare the lists to return is to slice off the elements that are already in common from the front of each of the sequences. Function all_equal could alternatively be implemented in a completely different way, less simple and obvious, but interesting: def all_equal(elements): return len(dict.fromkeys(elements)) == 1 or, equivalently and more concisely, in Python 2.4 only, def all_equal(elements): return len(set(elements)) == 1 Saying that all elements are equal is exactly the same as saying that the set of the elements has cardinality (length) one. In the variation using dict.fromkeys, we use a dict to represent the set, so that variation works in Python 2.3 as well as in 2.4. The variation using set is clearer, but it only works in Python 2.4. (You could also make it work in version 2.3, as well as Python 2.4, by using the standard Python library module sets). See Also Library Reference and Python in a Nutshell docs for modules os and itertools. Recipe 2.23. Reading an Unbuffered Character in a Cross-Platform Way Credit: Danny Yoo Problem Your application needs to read single characters, unbuffered, from standard input, and it needs to work on both Windows and Unix-like systems. Solution When we need a cross-platform solution, starting with platform-dependent ones, we need to wrap the different solutions so that they look the same: try: from msvcrt import getch except ImportError: ''' we're not on Windows, so we try the Unix-like approach ''' def getch( ): import sys, tty, termios fd = sys.stdin.fileno( ) old_settings = termios.tcgetattr(fd) try: tty.setraw(fd) ch = sys.stdin.read(1) finally: termios.tcsetattr(fd, termios.TCSADRAIN, old_settings) return ch Discussion On Windows, the standard Python library module msvcrt offers the handy getch function to read one character, unbuffered, from the keyboard, without echoing it to the screen. However, this module is not part of the standard Python library on Unix and Unix-like platforms, such as Linux and Mac OS X. On such platforms, we can get the same functionality with the tty and termios modules of the standard Python library (which, in turn, are not present on Windows). The key point is that in application-level code, we should never have to worry about such issues; rather, we should write our application code in platform-independent ways, counting on library functions to paper over the differences between platforms. The Python Standard Library fulfills that role admirably for most tasks, but not all, and the problem posed by this recipe is an example of one for which the Python Standard Library doesn't directly supply a cross-platform solution. When we can't find a ready-packaged cross-platform solution in the standard library, we should package it anyway as part of our own additional custom library. This recipe's Solution, besides solving the specific task of the recipe, also shows one good general way to go about such packaging. (Alternatively, you can test sys.platform, but I prefer the approach shown in this recipe.)Your own library module should try to import the standard library module it needs on a certain platform within a TRy clause and include a corresponding except ImportError clause that is triggered when the module is running on a different platform. In the body of that except clause, your own library module can apply whatever alternate approach will work on the different platform. In some rare cases, you may need more than two platform-dependent approaches, but most often you'll need one approach on Windows and only one other approach to cover all other platforms. This is because most non-Windows platforms today are generally Unix or Unix-like. See Also Library Reference and Python in a Nutshell docs for msvcrt, tty, and termios. Recipe 2.24. Counting Pages of PDF Documents on Mac OS X Credit: Dinu Gherman, Dan Wolfe Problem You're running on a reasonably recent version of Mac OS X (version 10.3 "Panther" or later), and you need to know the number of pages in a PDF document. Solution The PDF format and Python are both natively integrated with Mac OS X (10.3 or later), and this allows a rather simple solution: #!/usr/bin python import CoreGraphics def pageCount(pdfPath): "Return the number of pages for the PDF document at the given path." pdf = CoreGraphics.CGPDFDocumentCreateWithProvider( CoreGraphics.CGDataProviderCreateWithFilename(pdfPath) ) return pdf.getNumberOfPages( ) if _ _name_ _ == '_ _main_ _': import sys for path in sys.argv[1:]: print pageCount(path) Discussion A reasonable alternative to this recipe might be to use the PyObjC Python extension, which (among other wonders) lets Python code reuse all the power in the Foundation and AppKit frameworks that come with Mac OS X. Such a choice would let you write a Python script that is also able to run on older versions of Mac OS X, such as 10.2 Jaguar. However, relying on Mac OS X 10.3 or later ensures we can use the Python installation that is integrated as a part of the operating system, as well as such goodies as the CoreGraphics Python extension module (also part of Mac OS X "Panther") that lets your Python code reuse Apple's excellent Quartz graphics engine directly. See Also PyObjC is at http://pyobjc.sourceforge.net/; information on the CoreGraphics module is at http://www.macdevcenter.com/pub/a/mac/2004/03/19/core_graphics.html. Recipe 2.25. Changing File Attributes on Windows Credit: John Nielsen Problem You need to set the attributes of a file on Windows; for example, you may need to set the file as read-only, archived, and so on. Solution PyWin32's win32api module offers a function SetFileAttributes that makes this task quite simple: import win32con, win32api, os # create a file, just to show how to manipulate it thefile = 'test' f = open('test', 'w') f.close( ) # to make the file hidden...: win32api.SetFileAttributes(thefile, win32con.FILE_ATTRIBUTE_HIDDEN) # to make the file readonly: win32api.SetFileAttributes(thefile, win32con.FILE_ATTRIBUTE_READONLY) # to be able to delete the file we need to set it back to normal: win32api.SetFileAttributes(thefile, win32con.FILE_ATTRIBUTE_NORMAL) # and finally we remove the file we just made os.remove(thefile) Discussion One interesting use of win32api.SetFileAttributes is to enable a file's removal. Removing a file with os.remove can fail on Windows if the file's attributes are not normal. To get around this problem, you just need to use the Win32 call to SetFileAttributes to convert it to a normal file, as shown at the end of this recipe's Solution. Of course, this should be done with caution, since there may be a good reason the file is not "normal". The file should be removed only if you know what you're doing! See Also The documentation on the win32file module at http://ASPN.ActiveState.com/ASPN/Python/Reference/Products/ActivePython/PythonWin32Extensions/win32file.html. Recipe 2.26. Extracting Text from OpenOffice.org Documents Credit: Dirk Holtwick Problem You need to extract the text content (with or without the attending XML markup) from an OpenOffice.org document. Solution An OpenOffice.org document is just a zip file that aggregates XML documents according to a well-documented standard. To access our precious data, we don't even need to have OpenOffice.org installed: import zipfile, re rx_stripxml = re.compile("<[^>]*?>", re.DOTALL|re.MULTILINE) def convert_OO(filename, want_text=True): """ Convert an OpenOffice.org document to XML or text. """ zf = zipfile.ZipFile(filename, "r") data = zf.read("content.xml") zf.close( ) if want_text: data = " ".join(rx_stripxml.sub(" ", data).split( )) return data if _ _name_ _=="_ _main_ _": import sys if len(sys.argv)>1: for docname in sys.argv[1:]: print 'Text of', docname, ':' print convert_OO(docname) print 'XML of', docname, ':' print convert_OO(docname, want_text=False) else: print 'Call with paths to OO.o doc files to see Text and XML forms.' Discussion OpenOffice.org documents are zip files, and in addition to other contents, they always contain the file content.xml. This recipe's job, therefore, essentially boils down to just extracting this file. By default, the recipe then throws away XML tags with a simple regular expression, splits the result by whitespace, and joins it up again with a single blank to save space. Of course, we could use an XML parser to get information in a vastly richer and more structured way, but if all we need is the rough textual content, this fast, rough-and-ready approach may suffice. Specifically, the regular expression rx_stripxml matches any XML tag (opening or closing) from the leading < to the terminating >. Inside function convert_OO, in the statements guarded by if want_text, we use that regular expression to change every XML tag into a space, then normalize whitespace by splitting (i.e., calling the string method split, which splits on any sequence ofwhitespace), and rejoining (with " ".join, to use a single blank character as the joiner). Essentially, this split-and-rejoin process changes any sequence of whitespace into a single blank character. More advanced ways to extract all text from an XML document are shown in Recipe 12.3. See Also Library Reference docs on modules zipfile and re; OpenOffice.org's web site, http://www.openoffice.org/; Recipe 12.3. Recipe 2.27. Extracting Text from Microsoft Word Documents Credit: Simon Brunning, Pavel Kosina Problem You want to extract the text content from each Microsoft Word document in a directory tree on Windows into a corresponding text file. Solution With the PyWin32 extension, we can access Word itself, through COM, to perform the conversion: import fnmatch, os, sys, win32com.client wordapp = win32com.client.gencache.EnsureDispatch("Word.Application") try: for path, dirs, files in os.walk(sys.argv[1]): for filename in files: if not fnmatch.fnmatch(filename, '*.doc'): continue doc = os.path.abspath(os.path.join(path, filename)) print "processing %s" % doc wordapp.Documents.Open(doc) docastxt = doc[:-3] + 'txt' wordapp.ActiveDocument.SaveAs(docastxt, FileFormat=win32com.client.constants.wdFormatText) wordapp.ActiveDocument.Close( ) finally: # ensure Word is properly shut down even if we get an exception wordapp.Quit( ) Discussion A useful aspect of most Windows applications is that you can script them via COM, and the PyWin32 extension makes it fairly easy to perform COM scripting from Python. The extension enables you to write Python scripts to perform many kinds of Window tasks. The script in this recipe's Solution drives Microsoft Word to extract the text from every .doc file in a "directory" tree into a corresponding .txt text file. Using the os.walk function, we can access every subdirectory in a tree with a simple for statement, without recursion. With the fnmatch.fnmatch function, we can check a filename to determine whether it matches an appropriate wildcard, here '*.doc'. Once we have determined the name of a Word document file, we process that name with functions from os.path to turn it into a complete absolute path, and have Word open it, save it as text, and close it again. If you don't have Word, you may need to take a completely different approach. One possibility is to use OpenOffice.org, which is able to load Word documents. Another is to use a program specifically designed to read Word documents, such as Antiword, found at http://www.winfield.demon.nl/. However, we have not explored these alternative options.See Also Mark Hammond, Andy Robinson, Python Programming on Win32 (O'Reilly), for documentation on PyWin32; http://msdn.microsoft.com, for Microsoft's documentation of the object model of Microsoft Word; Library Reference and Python in a Nutshell sections on modules fnmatch and os.path, and function os.walk. Recipe 2.28. File Locking Using a Cross-Platform API Credit: Jonathan Feinberg, John Nielsen Problem You need to lock files in a program that runs on both Windows and Unix-like systems, but the Python Standard Library offers only platform-specific ways to lock files. Solution When the Python Standard Library itself doesn't offer a cross-platform solution, it's often possible to implement one ourselves: import os # needs win32all to work on Windows (NT, 2K, XP, _not_ /95 or /98) if os.name == 'nt': import win32con, win32file, pywintypes LOCK_EX = win32con.LOCKFILE_EXCLUSIVE_LOCK LOCK_SH = 0 # the default LOCK_NB = win32con.LOCKFILE_FAIL_IMMEDIATELY _ _overlapped = pywintypes.OVERLAPPED( ) def lock(file, flags): hfile = win32file._get_osfhandle(file.fileno( )) win32file.LockFileEx(hfile, flags, 0, 0xffff0000, _ _overlapped) def unlock(file): hfile = win32file._get_osfhandle(file.fileno( )) win32file.UnlockFileEx(hfile, 0, 0xffff0000, _ _overlapped) elif os.name == 'posix': from fcntl import LOCK_EX, LOCK_SH, LOCK_NB def lock(file, flags): fcntl.flock(file.fileno( ), flags) def unlock(file): fcntl.flock(file.fileno( ), fcntl.LOCK_UN) else: raise RuntimeError("PortaLocker only defined for nt and posix platforms") Discussion When multiple programs or threads have to access a shared file, it's wise to ensure that accesses are synchronized so that two processes don't try to modify the file contents at the same time. Failure to synchronize accesses could even corrupt the entire file in some cases. This recipe supplies two functions, lock and unlock, that request and release locks on a file, respectively. Using the portalocker.py module is a simple matter of calling the lock function and passing in the file and an argument specifying the kind of lock that is desired: Shared lock (default)This lock denies all processes, including the process that first locks the file, write access to the file. All processes can read the locked file. Exclusive lock This denies all other processes both read and write access to the file. Nonblocking lock When this value is specified, the function returns immediately if it is unable to acquire the requested lock. Otherwise, it waits. LOCK_NB can be ORed with either LOCK_SH or LOCK_EX by using Python's bitwise-or operator, the vertical bar (|). For example: import portalocker afile = open("somefile", "r+") portalocker.lock(afile, portalocker.LOCK_EX) The implementation of the lock and unlock functions is entirely different on different systems. On Unix-like systems (including Linux and Mac OS X), the recipe relies on functionality made available by the standard fcntl module. On Windows systems (NT, 2000, XPit doesn't work on old Win/95 and Win/98 platforms because they just don't have the needed oomph in the operating system!), the recipe uses the win32file module, part of the very popular PyWin32 package of Windows-specific extensions to Python, authored by Mark Hammond. But the important point is that, despite the differences in implementation, the functions (and the flags you can pass to the lock function) are made to behave in the same way across platforms. Such cross-platform packaging of differently implemented but equivalent functionality enables you to easily write cross-platform applications, which is one of Python's strengths. When you write a cross-platform program, it's nice if the functionality that your program uses is, in turn, encapsulated in a cross-platform way. For file locking in particular, it is especially helpful to Perl users, who are used to an essentially transparent lock system call across platforms. More generally, if os.name== just does not belong in application-level code. Such platform testing ideally should always be in the standard library or an application- independent module, as it is here. See Also Documentation on the fcntl module in the Library Reference; documentation on the win32file module at http://ASPN.ActiveState.com/ASPN/Python/Reference/Products/ActivePython/PythonWin32Extensions/win32file.html; Jonathan Feinberg's web site (http://MrFeinberg.com). Recipe 2.29. Versioning Filenames Credit: Robin Parmar, Martin Miller Problem You want to make a backup copy of a file, before you overwrite it, with the standard convention of appending a three-digit version number to the name of the old file. Solution We just need to code a function to perform the backup copy appropriately: def VersionFile(file_spec, vtype='copy'): import os, shutil if os.path.isfile(file_spec): # check the 'vtype' parameter if vtype not in ('copy', 'rename'): raise ValueError, 'Unknown vtype %r' % (vtype,) # Determine root filename so the extension doesn't get longer n, e = os.path.splitext(file_spec) # Is e a three-digits integer preceded by a dot? if len(e) == 4 and e[1:].isdigit( ): num = 1 + int(e[1:]) root = n else: num = 0 root = file_spec # Find next available file version for i in xrange(num, 1000): new_file = '%s.%03d' % (root, i) if not os.path.exists(new_file): if vtype == 'copy': shutil.copy(file_spec, new_file) else: os.rename(file_spec, new_file) return True raise RuntimeError, "Can't %s %r, all names taken"%(vtype,file_spec) return False if _ _name_ _ == '_ _main_ _': import os # create a dummy file 'test.txt' tfn = 'test.txt' open(tfn, 'w').close( ) # version it 3 times print VersionFile(tfn) # emits: True print VersionFile(tfn) # emits: True print VersionFile(tfn) # emits: True # remove all test.txt* files we just made for x in ('', '.000', '.001', '.002'): os.unlink(tfn + x) # show what happens when the file does not exist print VersionFile(tfn) # emits: False print VersionFile(tfn) # emits: False Discussion The purpose of the VersionFile function is to ensure that an existing file is copied (or renamed, as indicated by the optional second parameter) before you open it for writing or updating and therefore modify it. It is polite to make such backups of files before you mangle them (one functionality some people still pine for from the good old VMS operating system, which performed it automatically!). The actual copy or renaming is performed by shutil.copy and os.rename, respectively, so the only issue is which name to use as the target. A popular way to determine backups' names is versioning (i.e., appending to the filename a gradually incrementing number). This recipe determines the new name by first extracting the filename's root (just in case you call it with an already-versioned filename) and then successively appending to that root the further extensions .000, .001, and so on, until a name built in this manner does not correspond to any existing file. Then, and only then, is the name used as the target of a copy or renaming. Note that VersionFile is limited to 1,000 versions, so you should have an archive plan after that. The file must exist before it is first versionedyou cannot back up what does not yet exist. However, if the file doesn't exist, function VersionFile simply returns False (while it returns TRue if the file exists and has been successfully versioned), so you don't need to check before calling it! See Also Documentation for the os and shutil modules in the Library Reference and Python in a Nutshell. Recipe 2.30. Calculating CRC-64 Cyclic Redundancy Checks Credit: Gian Paolo Ciceri Problem You need to ensure the integrity of some data by computing the data's cyclic redundancy check (CRC), and you need to do so according to the CRC-64 specifications of the ISO-3309 standard. Solution The Python Standard Library does not include any implementation of CRC-64 (only one of CRC- 32 in function zlib.crc32), so we need to program it ourselves. Fortunately, Python can perform bitwise operations (masking, shifting, bitwise-and, bitwise-or, xor, etc.) just as well as, say, C (and, in fact, with just about the same syntax), so it's easy to transliterate a typical reference implementation of CRC-64 into a Python function as follows: # prepare two auxiliary tables tables (using a function, for speed), # then remove the function, since it's not needed any more: CRCTableh = [0] * 256 CRCTablel = [0] * 256 def _inittables(CRCTableh, CRCTablel, POLY64REVh, BIT_TOGGLE): for i in xrange(256): partl = i parth = 0L for j in xrange(8): rflag = partl & 1L partl >>= 1L if parth & 1: partl ^= BIT_TOGGLE parth >>= 1L if rflag: parth ^= POLY64REVh CRCTableh[i] = parth CRCTablel[i] = partl # first 32 bits of generator polynomial for CRC64 (the 32 lower bits are # assumed to be zero) and bit-toggle mask used in _inittables POLY64REVh = 0xd8000000L BIT_TOGGLE = 1L << 31L # run the function to prepare the tables _inittables(CRCTableh, CRCTablel, POLY64REVh, BIT_TOGGLE) # remove all names we don't need any more, including the function del _inittables, POLY64REVh, BIT_TOGGLE # this module exposes the following two functions: crc64, crc64digest def crc64(bytes, (crch, crcl)=(0,0)): for byte in bytes: shr = (crch & 0xFF) << 24 temp1h = crch >> 8L temp1l = (crcl >> 8L) | shr tableindex = (crcl ^ ord(byte)) & 0xFF crch = temp1h ^ CRCTableh[tableindex] crcl = temp1l ^ CRCTablel[tableindex] return crch, crcl def crc64digest(aString): return "%08X%08X" % (crc64(bytes)) if _ _name_ _ == '_ _main_ _': # a little test/demo, for when this module runs as main-script assert crc64("IHATEMATH") == (3822890454, 2600578513) assert crc64digest("IHATEMATH") == "E3DCADD69B01ADD1" print 'crc64: dumb test successful' Discussion Cyclic redundancy checks (CRCs) are a popular way to ensure that data (in particular, a file) has not been accidentally damaged. CRCs can readily detect accidental damage, but they are not intended to withstand inimical assault the way other cryptographically strong checksums are. CRCs can be computed much faster than other kinds of checksums, making them useful in those cases where the only damage we need to guard against is accidental damage, rather than deliberate adversarial tampering. Mathematically speaking, a CRC is computed as a polynomial over the bits of the data we're checksumming. In practice, as this recipe shows, most of the computation can be done once and for all and summarized in tables that, when properly indexed, give the contribution of each byte of input data to the result. So, after initialization (which we do with an auxiliary function because computation in Python is much faster when using a function's local variables than when using globals), actual CRC computation is quite fast. Both the computation of the tables and their use for CRC computation require a lot of bitwise operations, but, fortunately, Python's just as good at such operations as other languages such as C. (In fact, Python's syntax for the various bitwise operands is just about the same as C's.) The algorithm to compute the standard CRC-64 checksum is described in the ISO-3309 standard, and this recipe does nothing more than implement that algorithm. The generator polynomial is x64 + x4 + x3 + x + 1. (The "See Also" section within this recipe provides a reference for obtaining information about the computation.) We represent the 64-bit result as a pair of Python ints, holding the low and high 32-bit halves of the result. To allow the CRC to be computed incrementally, in those cases where the data comes in a little at a time, we let the caller of function crc64 optionally feed in the "initial value" for the (crch, crcl) pair, presumably obtained by calling crc64 on previous parts of the data. To compute the CRC in one gulp, the caller just needs to pass in the data (a string of bytes), since in this case, we initialize the result to (0, 0) by default. See Also W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery, Numerical Recipes in C, 2d ed. (Cambridge University Press), pp. 896ff. Chapter 3. Time and Money Introduction Recipe 3.1. Calculating Yesterday and Tomorrow Recipe 3.2. Finding Last Friday Recipe 3.3. Calculating Time Periods in a Date Range Recipe 3.4. Summing Durations of Songs Recipe 3.5. Calculating the Number of Weekdays Between Two Dates Recipe 3.6. Looking up Holidays Automatically Recipe 3.7. Fuzzy Parsing of Dates Recipe 3.8. Checking Whether Daylight Saving Time Is Currently in Effect Recipe 3.9. Converting Time Zones Recipe 3.10. Running a Command Repeatedly Recipe 3.11. Scheduling Commands Recipe 3.12. Doing Decimal Arithmetic Recipe 3.13. Formatting Decimals as Currency Recipe 3.14. Using Python as a Simple Adding Machine Recipe 3.15. Checking a Credit Card Checksum Recipe 3.16. Watching Foreign Exchange Rates Introduction Credit: Gustavo Niemeyer, Facundo Batista Today, last weekend, next year. These terms sound so common. You have probably wondered, at least once, about how deeply our lives are involved in the very idea of time. The concept of time surrounds us, and, as a consequence, it's also present in the vast majority of software projects. Even very simple programs may have to deal with timestamps, delays, timeouts, speed gauges, calendars, and so on. As befits a general-purpose language that is proud to come with "batteries included," Python's standard library offers solid support for these application needs, and more support yet comes from third-party modules and packages. Computing tasks involving money are another interesting topic that catches our attention because it's so closely related to our daily lives. Python 2.4 introduced support for decimal numbers (and you can retrofit that support into 2.3, see http://www.taniquetil.com.ar/facundo/bdvfiles/get_decimal.html), making Python a good option even for computations where you must avoid using binary floats, as ones involving money so often are. This chapter covers exactly these two topics, money and time. According to the old saying, maybe we should claim the chapter is really about a single topic, since after all, as everybody knowstime is money! The time Module Python Standard Library's time module lets Python applications access a good portion of the time-related functionality offered by the platform Python is running on. Your platform's documentation for the equivalent functions in the C library will therefore be useful, and some peculiarities of different platforms will affect Python as well. One of the most used functions from module time is the one that obtains the current timetime.time. This function's return value may be a little cryptic for the uninitiated: it's a floating-point number that corresponds to the number of seconds passed since a fixed instant called the epoch, which may change depending on your platform but is usually midnight of January 1, 1970. To check which epoch your platform uses, try, at any Python interactive interpreter prompt: >>> import time >>> print time.asctime(time.gmtime(0)) Notice we're passing 0 (meaning 0 seconds after the epoch) to the time.gmtime function. time.gmtime converts any timestamp (in seconds since the epoch) into a tuple that represents that precise instant of time in human terms, without applying any kind of time zone conversion (GMT stands for "Greenwich mean time", an old but colorful way to refer to what is now known as UTC, for "Coordinated Universal Time"). You can also pass a timestamp (in seconds since the epoch) to time.localtime, which applies the current local notion of time zone. It's important to understand the difference, since, if you have a timestamp that is already offset to represent a local time, passing it to the time.localtime function will not yield the expected resultunless you're so lucky that your local time zone happens to coincide with the UTC time zone, of course! Here is a way to unpack a tuple representing the current local time:year, month, mday, hour, minute, second, wday, yday = time.localtime( ) While valid, this code is not elegant, and it would certainly not be practical to use it often. This kind of construct may be completely avoided, since the tuples returned by the time functions let you access their elements via meaningful attribute names. Obtaining the current month then becomes a simple and elegant expression: time.localtime( ).tm_mon Note that we omitted passing any argument to localtime. When we call localtime, gmtime, or asctime without an argument, each of them conveniently defaults to using the current time. Two very useful functions in module time are strftime, which lets you build a string from a time tuple, and strptime, which goes the other way, parsing a string and producing a time tuple. Each of these two functions accepts a format string that lets you specify exactly what you want in the resulting string (or, respectively, what you expect from the string you're parsing) in excruciating detail. For all the formatting specifications that you can use in the format strings you pass to these functions, see http://docs.python.org/lib/module-time.html. One last important function in module time is the time.sleep function, which lets you introduce delays in Python programs. Even though this function's POSIX counterpart accepts only an integer parameter, the Python equivalent supports a float and allows sub-second delays to be achieved. For instance: for i in range(10): time.sleep(0.5) print "Tick!" This snippet will take about 5 seconds to execute, emitting Tick! approximately twice per second. Time and Date Objects While module time is quite useful, the Python Standard Library also includes the datetime module, which supplies types that provide better abstractions for the concepts of dates and timesnamely, the types time, date, and datetime. Constructing instances of those types is quite elegant: today = datetime.date.today( ) birthday = datetime.date(1977, 5, 4) #May 4 currenttime = datetime.datetime.now( ).time( ) lunchtime = datetime.time(12, 00) now = datetime.datetime.now( ) epoch = datetime.datetime(1970, 1, 1) meeting = datetime.datetime(2005, 8, 3, 15, 30) Further, as you'd expect, instances of these types offer comfortable information access and useful operations through their attributes and methods. The following statements create an instance of the date type, representing the current day, then obtain the same date in the next year, and finally print the result in a dotted format: today = datetime.date.today( ) next_year = today.replace(year=today.year+1).strftime("%Y.%m.%d") print next_yearNotice how the year was incremented, using the replace method. Assigning to the attributes of date and time instances may sound tempting, but these instances are immutable (which is a good thing, because it means we can use the instances as members in a set, or keys in a dictionary!), so new instances must be created instead of changing existing ones. Module datetime also provides basic support for time deltas (differences between instants of time; you can think of them as basically meaning durations in time), through the timedelta type. This type lets you change a given date by incrementing or decrementing the date by a given time slice, and it is also the result of taking the difference between times or dates. >>> import datetime >>> NewYearsDay = datetime.date(2005, 01, 01) >>> NewYearsEve = datetime.date(2004, 12, 31) >>> oneday = NewYearsDay - NewYearsEve >>> print oneday 1 day, 0:00:00 >>> A timedelta instance is internally represented by days, seconds, and microseconds, but you can construct timedelta instances by passing any of these arguments and also other multipliers, like minutes, hours and weeks. Support for other kinds of deltas, like months, and years, is not availableon purpose, since their meanings, and operation results, are debatable. (This feature is, however, offered by the third-party dateutil packagesee https://moin.conectiva.com.br/DateUtil.) datetime can be described as a prudent or cautious design. The decision of not implementing doubtful tasks, and tasks that may need many different implementations in different systems, reflects the strategy used to develop all of the module. This way, the module offers good interfaces for most use cases, and, even more importantly, a strong and coherent base for third- party modules to build upon. Another area where this cautious design strategy for datetime shows starkly is the module's time zone support. Even though datetime offers nice ways to query and set time zone information, they're not really useful without an external source to provide nonabstract subclasses of the tzinfo type. At least two third-party packages provide time zone support for datetime: dateutil, mentioned previously, and pyTZ, available at http://sourceforge.net/projects/pytz/. Decimal decimal is a Python Standard Library module, new in Python 2.4, which finally brings decimal arithmetic to Python. Thanks to decimal, we now have a decimal numeric data type, with bounded precision and floating point. Let's look at each of these three little phrases in more detail: Decimal numeric data type The number is not stored in binary, but rather, as a sequence of decimal digits. With bounded precision The number of digits each number stores is fixed. (It is a fixed parameter of each decimalnumber object, but different decimal number objects can be set to use different numbers of digits.) Floating point The decimal point does not have a fixed place. (To put it another way: while the number has a fixed amount of digits in total, it does not have a fixed amount of digits after the decimal point. If it did, it would be a fixed-point, rather than floating-point, numeric data type). Such a data type has many uses (the big use case is as the basis for money computations), particularly because decimal.Decimal offers many other advantages over standard binary float. The main advantage is that all of the decimal numbers that the user can enter (which is to say, all the decimal numbers with a finite number of digits) can be represented exactly (in contrast, some of those numbers do not have an exact representation in binary floating point): >>> import decimal >>> 1.1 1.1000000000000001 >>> 2.3 2.2999999999999998 >>> decimal.Decimal("1.1") Decimal("1.1") >>> decimal.Decimal("2.3") Decimal("2.3") The exactness of the representation carries over into arithmetic. In binary floating point, for example: >>> 0.1 + 0.1 + 0.1 - 0.3 5.5511151231257827e-17 Such differences are very close to zero, and yet they prevent reliable equality testing; moreover, even tiny differences can accumulate. For this reason, decimal should be preferred to binary floats in accounting applications that have strict equality requirements: >>> d1 = decimal.Decimal("0.1") >>> d3 = decimal.Decimal("0.3") >>> d1 + d1 + d1 - d3 Decimal("0.0") decimal.Decimal instances can be constructed from integers, strings, or tuples. To create a decimal.Decimal from a float, first convert the float to a string. This necessary step serves as an explicit reminder of the details of the conversion, including representation error. Decimal numbers include special values such as NaN (which stands for "not a number"), positive and negative Infinity, and -0. Once constructed, a decimal.Decimal object is immutable, just like any other number in Python. The decimal module essentially implements the rules of arithmetic that are taught in school. Up to a given working precision, exact, unrounded results are given whenever possible: >>> 0.9 / 10 0.089999999999999997 >>> decimal.Decimal("0.9") / decimal.Decimal(10) Decimal("0.09")Where the number of digits in a result exceeds the working precision, the number is rounded according to the current rounding method. Several rounding methods are available; the default is round-half-even. The decimal module incorporates the notion of significant digits, so that, for example, 1.30+1.20 is 2.50. The trailing zero is kept to indicate significance. This is the usual representation for monetary applications. For multiplication, the "schoolbook" approach uses all the figures in the multiplicands: >>> decimal.Decimal("1.3") * decimal.Decimal("1.2") Decimal("1.56") >>> decimal.Decimal("1.30") * decimal.Decimal("1.20") Decimal("1.5600") In addition to the standard numeric properties that decimal objects share with other built-in number types, such as float and int, decimal objects also have several specialized methods. Check the docs for all of the methods, with details and examples. The decimal data type works within a context, where some configuration aspects are set. Each thread has its own current context (having a separate context per thread means that each thread may make changes without interfering with other threads); the current thread's current context is accessed or changed using functions getcontext and setcontext from the decimal module. Unlike hardware-based binary floating point, the precision of the decimal module can be set by users (defaulting to 28 places). It can be set to be as large as needed for a given problem: >>> decimal.getcontext( ).prec = 6 # set the precision to 6... >>> decimal.Decimal(1) / decimal.Decimal(7) Decimal("0.142857") >>> decimal.getcontext( ).prec = 60 # ...and to 60 digits >>> decimal.Decimal(1) / decimal.Decimal(7) Decimal("0.142857142857142857142857142857142857142857142857142857142857") Not everything in decimal can be as simple and elementary as shown so far, of course. Essentially, decimal implements the standards for general decimal arithmetic which you can study in detail at http://www2.hursley.ibm.com/decimal/. In particular, this means that decimal supports the concept of signals. Signals represent abnormal conditions arising from computations (e.g., 1/0, 0/0, Infinity/Infinity). Depending on the needs of each specific application, signals may be ignored, considered as informational, or treated as exceptions. For each signal, there is a flag and a trap enabler. When a signal is encountered, its flag is incremented from zero, and then, if the trap enabler is set to one, an exception is raised. This gives programmers a great deal of power and flexibility in configuring decimal to meet their exact needs. Given all of these advantages for decimal, why would someone want to stick with float? Indeed, is there any reason why Python (like just about every other widespread language, with Cobol and Rexx the two major exceptions that easily come to mind) originally adopted floating- point binary numbers as its default (or only) noninteger data type? Of coursemany reasons can be provided, and they're all spelled speed! Consider: $ python -mtimeit -s'from decimal import Decimal as D' 'D("1.2")+D("3.4")' 10000 loops, best of 3: 191 usec per loop $ python -mtimeit -s'from decimal import Decimal as D' '1.2+3.4' 1000000 loops, best of 3: 0.339 usec per loop This basically translates to: on this machine (an old Athlon 1.2 GHz PC running Linux), Python can perform almost 3 million sums per second on floats (using the PC's arithmetic hardware),but only a bit more than 5 thousand sums per second on Decimals (all done in software and with all the niceties shown previously). Essentially, if your application must sum many tens of millions of noninteger numbers, you had better stick with float! When an average machine was a thousand times slower than it is today (and it wasn't all that long ago!), such limitations hit even applications doing relatively small amounts of computation, if the applications ran on reasonably cheap machines (again, we see time and money both playing a role!). Rexx and Cobol were born on mainframes that were not quite as fast as today's cheapest PCs but thousands of times more expensive. Purchasers of such mainframes could afford nice and friendly decimal arithmetic, but most other languages, born on more reasonably priced machines (or meant for computationally intensive tasks), just couldn't. Fortunately, relatively few applications actually need to perform so much arithmetic on non- integers as to give any observable performance problems on today's typical machines. Thus, today, most applications can actually take advantage of decimal's many beneficial aspects, including applications that must continue to use Python 2.3, even though decimal is in the Python Standard Library only since version 2.4. To learn how you can easily integrate decimal into Python 2.3, see http://www.taniquetil.com.ar/facundo/bdvfiles/get_decimal.html. Recipe 3.1. Calculating Yesterday and Tomorrow Credit: Andrea Cavalcanti Problem You want to get today's date, then calculate yesterday's or tomorrow's. Solution Whenever you have to deal with a "change" or "difference" in time, think timedelta: import datetime today = datetime.date.today( ) yesterday = today - datetime.timedelta(days=1) tomorrow = today + datetime.timedelta(days=1) print yesterday, today, tomorrow #emits: 2004-11-17 2004-11-18 2004-11-19 Discussion This recipe's Problem has been a fairly frequent question on Python mailing lists since the datetime module arrived. When first confronted with this task, it's quite common for people to try to code it as yesterday = today - 1, which gives a TypeError: unsupported operand type(s) for -: 'datetime.date' and 'int'. Some people have called this a bug, implying that Python should guess what they mean. However, one of the guiding principles that gives Python its simplicity and power is: "in the face of ambiguity, refuse the temptation to guess." Trying to guess would clutter datetime with heuristics meant to guess that you "really meant 1 day", rather than 1 second (which timedelta also supports), or 1 year. Rather than trying to guess what you mean, Python, as usual, expects you to make your meaning explicit. If you want to subtract a time difference of one day, you code that explicitly. If, instead, you want to add a time difference of one second, you can use timedelta with a datetime.datetime object, and then you code the operation using exactly the same syntax. This way, for each task you might want to perform, there's only one obvious way of doing it. This approach also allows a fair amount of flexibility, without added complexity. Consider the following interactive snippet: >>> anniversary = today + datetime.timedelta(days=365) # add 1 year >>> print anniversary 2005-11-18 >>> t = datetime.datetime.today( ) # get right now >>> t datetime.datetime(2004, 11, 19, 10, 12, 43, 801000) >>> t2 = t + datetime.timedelta(seconds=1) # add 1 second >>> t2 datetime.datetime(2004, 11, 19, 10, 12, 44, 801000) >>> t3 = t + datetime.timedelta(seconds=3600) # add 1 hour >>> t3datetime.datetime(2004, 11, 19, 11, 12, 43, 801000) Keep in mind that, if you want fancier control over date and time arithmetic, third-party packages, such as dateutil (which works together with the built-in datetime) and the classic mx.DateTime, are available. For example: from dateutil import relativedelta nextweek = today + relativedelta.relativedelta(weeks=1) print nextweek #emits: 2004-11-25 However, "always do the simplest thing that can possibly work." For simple, straightforward tasks such as the ones in this recipe, datetime.timedelta works just fine. See Also dateutil documentation at https://moin.conectiva.com.br/DateUtil?action=highlight&value= DateUtil, and datetime documentation in the Library Reference. mx.DateTime can be found at http://www.egenix.com/files/python/mxDateTime.html. mx.DateTime can be found at http://www.egenix.com/files/python/mxDateTime.html. Recipe 3.2. Finding Last Friday Credit: Kent Johnson, Danny Yoo, Jonathan Gennick, Michael Wener Problem You want to find the date of last Friday (or today, if today is Friday) and print it in a specified format. Solution You can use the datetime module from Python's standard library to easily achieve this: import datetime, calendar lastFriday = datetime.date.today( ) oneday = datetime.timedelta(days=1) while lastFriday.weekday( ) != calendar.FRIDAY: lastFriday -= oneday print lastFriday.strftime('%A, %d-%b-%Y') # emits, e.g.: Friday, 10-Dec-2004 Discussion The handy little snippet of code in this recipe lets us find a previous weekday and print the properly formatted date, regardless of whether that weekday is in the same month, or even the same year. In this example, we're looking for the last Friday (or today, if today is Friday). Friday's integer representation is 4, but to avoid depending on this "magical number," we just import the Python Standard Library calendar module and rely instead on its calendar.FRIDAY attribute (which, sure enough, is the number 4). We set a variable called lastFriday to today's date and work backward until we have reached a date with the desired weekday value of 4. Once we have the date we desire, formatting the date in any way we like is easily achieved with the "string formatting" method strftime of the datetime.date class. An alternative, slightly more terse solution uses the built-in constant datetime.date.resolution instead of explicitly building the datetime.timedelta instance to represent one day's duration: import datetime, calendar lastFriday = datetime.date.today( ) while lastFriday.weekday( ) != calendar.FRIDAY: lastFriday -= datetime.date.resolution print lastFriday.strftime('%d-%b-%Y') The datetime.date.resolution class attribute has exactly the same value as the oneday variable in the recipe's Solutionthe time interval of one day. However, resolution can trip you up. The value of the class attribute resolution varies among various classes of the datetime modulefor the date class it's timedelta(days=1), but for the time and datetime classes , it's timedelta(microseconds=1). You could mix-and-match (e.g., add datetime.date.resolution to a datetime.datetime instance), but it's easy to get confused doing so. The version in this recipe's Solution, using the explicitly named and defined oneday variable, is just as general,more explicit, and less confusing. Thus, all in all, that version is more Pythonic (which is why it's presented as the "official" one!). A more important enhancement is that we don't really need to loop, decrementing a date by one at each step through the loop: we can, in fact, get to the desired target in one fell swoop, computing the number of days to subtract thanks to the wonders of modular arithmetic: import datetime, calendar today = datetime.date.today( ) targetDay = calendar.FRIDAY thisDay = today.weekday( ) deltaToTarget = (thisDay - targetDay) % 7 lastFriday = today - datetime.timedelta(days=deltaToTarget) print lastFriday.strftime('%d-%b-%Y') If you don't follow why this works, you may want to brush up on modular arithmetic, for example at http://www.cut-the-knot.org/blue/Modulo.shtml. Use the approach you find clearest, without worrying about performance. Remember Hoare's dictum (often misattributed to Knuth, who was in fact quoting Hoare): "premature optimization is the root of all evil in programming." Let's see why thinking of optimization would be premature here. Net of the common parts (computing today's date, and formatting and emitting the result) on a four-year-old PC, with Linux and Python 2.4, the slowest approach (the one chosen for presentation as the "Solution" because it's probably the clearest and most obvious one) takes 18.4 microseconds; the fastest approach (the one avoiding the loop, with some further tweaks to really get pedal to the metal ) takes 10.1 microseconds. You're not going to compute last Friday's date often enough, in your life, to be able to tell the difference at 8 microseconds a pop (much less if you use recent hardware rather than a box that's four years old). If you consider the time needed to compute today's date and to format the result, you need to add 37 microseconds to each timing, even net of the I/O time for the print statement; so, the range of performance is roughly between 55 microseconds for the slowest and clearest form, and 47 microseconds for the fastest and tersest oneclearly not worth worrying about. See Also datetime module and strftime documentation in the Library Reference (currently at http://www.python.org/doc/lib/module-datetime.html and http://www.python.org/doc/current/lib/node208.html). Recipe 3.3. Calculating Time Periods in a Date Range Credit: Andrea Cavalcanti Problem Given two dates, you want to calculate the number of weeks between them. Solution Once again, the standard datetime and third-party dateutil modules (particularly dateutil's rrule.count method) come in quite handy. After importing the appropriate modules, it's a really simple job: from dateutil import rrule import datetime def weeks_between(start_date, end_date): weeks = rrule.rrule(rrule.WEEKLY, dtstart=start_date, until=end_date) return weeks.count( ) Discussion Function weeks_between takes the starting and ending dates as arguments, instantiates a rule to recur weekly between them, and returns the result of the rule's count methodfaster to code than to describe. This method will return only an integer (it won't return "half" weeks). For example, eight days is considered two weeks. It's easy to code a test for this: if _ _name_ _=='_ _main_ _': starts = [datetime.date(2005, 01, 04), datetime.date(2005, 01, 03)] end = datetime.date(2005, 01, 10) for s in starts: days = rrule.rrule(rrule.DAILY, dtstart=s, until=end).count( ) print "%d days shows as %d weeks "% (days, weeks_between(s, end)) This test emits the following output: 7 days shows as 1 weeks 8 days shows as 2 weeks It's not necessary to give a name to a recurrence rule, if you don't want tochanging the function's body, for example, to the single statement: return rrule.rrule(rrule.WEEKLY, dtstart=start_date, until=end_date).count( ) works just as well. I prefer to name recurrence rules because (frankly) I still find them a bit weird, even though they're so incredibly useful I doubt I could do without them!See Also Refer to the dateutil module's documentation available at https://moin.conectiva.com.br/DateUtil?action=highlight&value=DateUtil, datetime documentation in the Library Reference. Recipe 3.4. Summing Durations of Songs Credit: Anna Martelli Ravenscroft Problem You want to find out the total duration of a playlist of songs. Solution Use the datetime standard module and the built-in function sum to handle this task: import datetime def totaltimer(times): td = datetime.timedelta(0) # initial value of sum (must be a timedelta) duration = sum([ datetime.timedelta(minutes=m, seconds=s) for m, s in times], td) return duration if _ _name_ _== '_ _main_ _': # test when module run as main script times1 = [(2, 36), # list containing tuples (minutes, seconds) (3, 35), (3, 45),] times2 = [(3, 0), (5, 13), (4, 12), (1, 10),] assert totaltimer(times1) == datetime.timedelta(0, 596) assert totaltimer(times2) == datetime.timedelta(0, 815) print ("Tests passed.\n" "First test total: %s\n" "Second test total: %s" % ( totaltimer(times1), totaltimer(times2))) Discussion I have a large playlist of songs I listen to during workouts. I wanted to create a select list but wanted to know the total duration of the selected songs, without having to create the new playlist first. I wrote this little script to handle the task. A datetime.timedelta is normally what's returned when calculating the difference between two datetime objects. However, you can create your own timedelta instance to represent any given duration of time (while other classes of the datetime module, such as class datetime, have instances that represent a point in time). Here, we need to sum durations, so, clearly, it's exactly timedelta that we need. datetime.timedelta takes a variety of optional arguments: days, seconds, microseconds, milliseconds, minutes, hours, weeks. So, to create an instance, you really should pass named arguments when you call the class to avoid confusion. If you simply call datetime.timedelta(m, n), without naming the arguments, the class uses positional notation and treats m and n as days and seconds, which produces really strange results. (I found this out the hard way . . . a gooddemonstration of the need to test!) To use the built-in function sum on a list of objects such as timedeltas, you have to pass to sum a second argument to use as the initial valueotherwise, the default initial value is 0, integer zero, and you get an error as soon as you try to sum a timedelta with that int. All objects in the iterable that you pass as sum's first argument should be able to support numeric addition. (Strings are specifically disallowed, but, take my earnest advice: don't use sum for concatenating a lot of lists either!) In Python 2.4, instead of a list comprehension for sum's first argument, we could use a generator expression by replacing the square brackets, [ and ], with parentheses, ( and )which might be handy if you're trying to handle a playlist of several thousand songs. For the test case, I manually created a list of tuples with the durations of the songs in minutes and seconds. The script could be enhanced to parse the times in different formats (such as mm:ss) or to read the information from a file or directly from your music library. See Also Library Reference on sum and datetime. Recipe 3.5. Calculating the Number of Weekdays Between Two Dates Credit: Anna Martelli Ravenscroft Problem You want to calculate the number of weekdays (working days), as opposed to calendar days, that fall between two dates. Solution Since weekends and other "days off" vary by country, by region, even sometimes within a single company, there is no built-in way to perform this task. However, using dateutil along with datetime objects, it's reasonably simple to code a solution: from dateutil import rrule import datetime def workdays(start, end, holidays=0, days_off=None): if days_off is None: days_off = 5, 6 # default to: saturdays and sundays workdays = [x for x in range(7) if x not in days_off] days = rrule.rrule(rrule.DAILY, dtstart=start, until=end, byweekday=workdays) return days.count( ) - holidays if _ _name_ _ == '_ _main_ _': # test when run as main script testdates = [ (datetime.date(2004, 9, 1), datetime.date(2004, 11, 14), 2), (datetime.date(2003, 2, 28), datetime.date(2003, 3, 3), 1), ] def test(testdates, days_off=None): for s, e, h in testdates: print 'total workdays from %s to %s is %s with %s holidays' % ( s, e, workdays(s, e, h, days_off), h) test(testdates) test(testdates, days_off=[6]) Discussion This project was my very first one in Python: I needed to know the number of actual days in training of our trainees, given a start date and end date (inclusive). This problem was a bit trickier back in Python 2.2; today, the datetime module and the dateutil third-party package make the problem much simpler to solve. Function workdays starts by assigning a reasonable default value to variable days_off (unless an explicit value was passed for it as an argument), which is a sequence of the weekday numbers of our normal days off. In my company, weekly days off varied among individuals but were usually fewer than the workdays, so it was easier to track and modify the days off rather than the workdays. I made this an argument to the function so that I can easily pass a different value for days_off if and when I have different needs. Then, the function uses a list comprehension to create a list of actual weekly workdays, which are all weekdays not in days_off. Now the functionis ready to do its calculations. The workhorse in this recipe is an instance, named days, of dateutil's rrule (recurrence rule) class. Class rrule may be instantiated with various parameters to produce a rule object. In this example, I pass a frequency (rrule.DAILY), a beginning date and an ending dateboth of which must be datetime.date objectsand which weekdays to include (workdays). Then, I simply call method days.count to count the number of occurrences generated by the rule. (See Recipe 3.3 for other uses for the count method of rrule.) You can easily set your own definition of weekend: just pass as days_off whichever values you need. In this recipe, the default value is set to the standard U.S. weekend of Saturday and Sunday. However, if your company normally works a four-day week, say, Tuesday through Friday, you would pass days_off=(5, 6, 0). Just be sure to pass the days_off value as an iterable, such as a list or tuple, even if, as in the second test, you only have a single day in that container. A simple but useful enhancement might be to automatically check whether your start and end dates are weekends (for weekend-shift workers), and use an if/else to handle the weekend shifts, with appropriate changes to days_off. Further enhancements would be to add the ability to enter sick days, or to perform a call to an automatic holiday lookup function, rather than passing the number of holidays directly, as I do in this recipe. See Recipe 3.6 for a simple implementation of a holidays list for this purpose. See Also Refer to the dateutil documentation, which is available at https://moin.conectiva.com.br/DateUtil?action=highlight&value=DateUtil, datetime documentation in the Library Reference; Recipe 3.3 for another use of rrule.count; Recipe 3.6 for automatic holiday lookup. Recipe 3.6. Looking up Holidays Automatically Credit: Anna Martelli Ravenscroft, Alex Martelli Problem Holidays vary by country, by region, even by union within the same company. You want an automatic way to determine the number of holidays that fall between two given dates. Solution Between two dates, there may be movable holidays, such as Easter and Labor Day (U.S.); holidays that are based on Easter, such as Boxing Day; holidays with a fixed date, such as Christmas; holidays that your company has designated (the CEO's birthday). You can deal with all of them using datetime and the third-party module dateutil. A very flexible architecture is to factor out the various possibilities into separate functions to be called as appropriate: import datetime from dateutil import rrule, easter try: set except NameError: from sets import Set as set def all_easter(start, end): # return the list of Easter dates within start..end easters = [easter.easter(y) for y in xrange(start.year, end.year+1)] return [d for d in easters if start<=d<=end] def all_boxing(start, end): # return the list of Boxing Day dates within start..end one_day = datetime.timedelta(days=1) boxings = [easter.easter(y)+one_day for y in xrange(start.year, end.year+1)] return [d for d in boxings if start<=d<=end] def all_christmas(start, end): # return the list of Christmas Day dates within start..end christmases = [datetime.date(y, 12, 25) for y in xrange(start.year, end.year+1)] return [d for d in christmases if start<=d<=end] def all_labor(start, end): # return the list of Labor Day dates within start..end labors = rrule.rrule(rrule.YEARLY, bymonth=9, byweekday=rrule.MO(1), dtstart=start, until=end) return [d.date( ) for d in labors] # no need to test for in-between here def read_holidays(start, end, holidays_file='holidays.txt'): # return the list of dates from holidays_file within start..end try: holidays_file = open(holidays_file) except IOError, err: print 'cannot read holidays (%r):' % (holidays_file,), err return [ ] holidays = [ ] for line in holidays_file: # skip blank lines and comments if line.isspace( ) or line.startswith('#'): continue # try to parse the format: YYYY, M, D try: y, m, d = [int(x.strip( )) for x in line.split(',')] date = datetime.date(y, m, d) except ValueError: # diagnose invalid line and just go on print "Invalid line %r in holidays file %r" % ( line, holidays_file) continue if start<=date<=end: holidays.append(date) holidays_file.close( ) return holidays holidays_by_country = { # map each country code to a sequence of functions 'US': (all_easter, all_christmas, all_labor), 'IT': (all_easter, all_boxing, all_christmas), } def holidays(cc, start, end, holidays_file='holidays.txt'): # read applicable holidays from the file all_holidays = read_holidays(start, end, holidays_file) # add all holidays computed by applicable functions functions = holidays_by_country.get(cc, ( )) for function in functions: all_holidays += function(start, end) # eliminate duplicates all_holidays = list(set(all_holidays)) # uncomment the following 2 lines to return a sorted list: # all_holidays.sort( ) # return all_holidays return len(all_holidays) # comment this out if returning list if _ _name_ _ == '_ _main_ _': test_file = open('test_holidays.txt', 'w') test_file.write('2004, 9, 6\n') test_file.close( ) testdates = [ (datetime.date(2004, 8, 1), datetime.date(2004, 11, 14)), (datetime.date(2003, 2, 28), datetime.date(2003, 5, 30)), (datetime.date(2004, 2, 28), datetime.date(2004, 5, 30)), ] def test(cc, testdates, expected): for (s, e), expect in zip(testdates, expected): print 'total holidays in %s from %s to %s is %d (exp %d)' % ( cc, s, e, holidays(cc, s, e, test_file.name), expect) print test('US', testdates, (1,1,1) ) test('IT', testdates, (1,2,2) ) import os os.remove(test_file.name) Discussion In one company I worked for, there were three different unions, and holidays varied among the unions by contract. In addition, we had to track any snow days or other release days in the same way as "official" holidays. To deal with all the potential variations in holidays, it's easiest to factorout the calculation of standard holidays into their own functions, as we did in the preceding example for all_easter, all_labor, and so on. Examples of different types of calculations are provided so it's easy to roll your own as needed. Although half-open intervals (with the lower bound included but the upper one excluded) are the norm in Python (and for good reasons, since they're arithmetically more malleable and tend to induce fewer bugs in your computations!), this recipe deals with closed intervals instead (both lower and upper bounds included). Unfortunately, that's how specifications in terms of date intervals tend to be given, and dateutil also works that way, so the choice was essentially obvious. Each function is responsible for ensuring that it only returns results that meet our criteria: lists of datetime.date instances that lie between the dates (inclusive) passed to the function. For example, in all_labor, we coerce the datetime.datetime results returned by dateutil's rrule into datetime.date instances with the date method. A company may choose to set a specific date as a holiday (such as a snow day) "just this once," and a text file may be used to hold such unique instances. In our example, the read_holidays function handles the task of reading and processing a text file, with one date per line, each in the format year, month, day. You could also choose to refactor this function to use a "fuzzy" date parser, as shown in Recipe 3.7. If you need to look up holidays many times within a single run of your program, you may apply the optimization of reading and parsing the text file just once, then using the list of dates parsed from its contents each time that data is needed. However, "premature optimization is the root of all evil in programming," as Knuth said, quoting Hoare: by avoiding even this "obvious" optimization, we gain clarity and flexibility. Imagine these functions being used in an interactive environment, where the text file containing holidays may be edited between one computation and the next: by rereading the file each time, there is no need for any special check about whether the file was changed since you last read it! Since countries often celebrate different holidays, the recipe provides a rudimentary holidays_by_country dictionary. You can consult plenty of web sites that list holidays by country to flesh out the dictionary for your needs. The important part is that this dictionary allows a different group of holidays-generating functions to be called, depending on which country code is passed to the holidays function. If your company has multiple unions, you could easily create a union-based dictionary, passing the union-code instead of (or for multinationals, in addition to) a country code to holidays. The holidays function calls the appropriate functions (including, unconditionally, read_holidays), concatenates the results, eliminates duplicates, and returns the length of the list. If you prefer, of course, you can return the list instead, by simply uncommenting two lines as indicated in the code. See Also Recipe 3.7 for fuzzy parsing; dateutil documentation at https://moin.conectiva.com.br/DateUtil?action=highlight&value=DateUtil, datetime documentation in the Library Reference. Recipe 3.7. Fuzzy Parsing of Dates Credit: Andrea Cavalcanti Problem Your program needs to read and accept dates that don't conform to the datetime standard format of "yyyy, mm, dd". Solution The third-party dateutil.parser module provides a simple answer: import datetime import dateutil.parser def tryparse(date): # dateutil.parser needs a string argument: let's make one from our # `date' argument, according to a few reasonable conventions...: kwargs = { } # assume no named-args if isinstance(date, (tuple, list)): date = ' '.join([str(x) for x in date]) # join up sequences elif isinstance(date, int): date = str(date) # stringify integers elif isinstance(date, dict): kwargs = date # accept named-args dicts date = kwargs.pop('date') # with a 'date' str try: try: parsedate = dateutil.parser.parse(date, **kwargs) print 'Sharp %r -> %s' % (date, parsedate) except ValueError: parsedate = dateutil.parser.parse(date, fuzzy=True, **kwargs) print 'Fuzzy %r -> %s' % (date, parsedate) except Exception, err: print 'Try as I may, I cannot parse %r (%s)' % (date, err) if _ _name_ _ == "_ _main_ _": tests = ( "January 3, 2003", # a string (5, "Oct", 55), # a tuple "Thursday, November 18", # longer string without year "7/24/04", # a string with slashes "24-7-2004", # European-format string {'date':"5-10-1955", "dayfirst":True}, # a dict including the kwarg "5-10-1955", # dayfirst, no kwarg 19950317, # not a string "11AM on the 11th day of 11th month, in the year of our Lord 1945", ) for test in tests: # testing date formats tryparse(test) # try to parse Discussiondateutil.parser's parse function works on a variety of date formats. This recipe demonstrates a few of them. The parser can handle English-language month-names and two- or four-digit years (with some constraints). When you call parse without named arguments, its default is to first try parsing the string argument in the following order: mm-dd-yy. If that does not make logical sense, as, for example, it doesn't for the '24-7-2004' string in the recipe, parse then tries dd-mm-yy. Lastly, it tries yy-mm-dd. If a "keyword" such as dayfirst or yearfirst is passed (as we do in one test), parse attempts to parse based on that keyword. The recipe tests define a few edge cases that a date parser might encounter, such as trying to pass the date as a tuple, an integer (ISO-formatted without spaces), and even a phrase. To allow testing of the keyword arguments, the tryparse function in the recipe also accepts a dictionary argument, expecting, in this case, to find in it the value of the string to be parsed in correspondence to key 'date', and passing the rest on to dateutil's parser as keyword arguments. dateutil's parser can provide a pretty good level of "fuzzy" parsing, given some hints to let it know which piece is, for example, the hour (such as the AM in the test phrase in this recipe). For production code, you should avoid relying on fuzzy parsing, and either do some kind of preprocessing, or at least provide some kind of mechanism for checking the accuracy of the parsed date. See Also For more on date-parsing algorithms, see dateutil documentation at https://moin.conectiva.com.br/DateUtil?action=highlight&value=DateUtil; for date handling, see the datetime documentation in the Library Reference. Recipe 3.8. Checking Whether Daylight Saving Time Is Currently in Effect Credit: Doug Fort Problem You want to know whether daylight saving time is in effect in your local time zone today. Solution It's a natural temptation to check time.daylight for this purpose, but that doesn't work. Instead you need: import time def is_dst( ): return bool(time.localtime( ).tm_isdst) Discussion In my location (as in most others nowadays), time.daylight is always 1 because time.daylight means that this time zone has daylight saving time (DST) at some time during the year, whether or not DST is in effect today. The very last item in the pseudo-tuple you get by calling time.localtime, on the other hand, is 1 only when DST is currently in effect, otherwise it's 0which, in my experience, is exactly the information one usually needs to check. This recipe wraps this check into a function, calling built- in type bool to ensure the result is an elegant true or False rather than a rougher 1 or 0optional refinements, but nice ones, I think. You could alternatively access the relevant item as time.localtime( )[-1], but using attribute-access syntax with the tm_isdst attribute name is more readable. See Also Library Reference and Python in a Nutshell about module time. Recipe 3.9. Converting Time Zones Credit: Gustavo Niemeyer Problem You are in Spain and want to get the correct local (Spanish) time for an event in China. Solution Time zone support for datetime is available in the third-party dateutil package. Here's one way to set the local time zone, then print the current time to check that it worked properly: from dateutil import tz import datetime posixstr = "CET-1CEST-2,M3.5.0/02:00,M10.5.0/03:00" spaintz = tz.tzstr(posixstr) print datetime.datetime.now(spaintz).ctime( ) Conversion between different time zones is also possible, and often necessary in our expanding world. For instance, let's find out when the next Olympic Games will start, according to a Spanish clock: chinatz = tz.tzoffset("China", 60*60*8) olympicgames = datetime.datetime(2008, 8, 8, 20, 0, tzinfo=chinatz) print olympicgames.astimezone(spaintz) Discussion The cryptic string named posixstr is a POSIX-style representation for the time zone currently being used in Spain. This string provides the standard and daylight saving time zone names (CST and CEST), their offsets (UTC+1 and UTC+2), and the day and hour when DST starts and ends (the last Sunday of March at 2 a.m., and the last Sunday of October at 3 a.m., respectively). We may check the DST zone bounds to ensure they are correct: assert spaintz.tzname(datetime.datetime(2004, 03, 28, 1, 59)) == "CET" assert spaintz.tzname(datetime.datetime(2004, 03, 28, 2, 00)) == "CEST" assert spaintz.tzname(datetime.datetime(2004, 10, 31, 1, 59)) == "CEST" assert spaintz.tzname(datetime.datetime(2004, 10, 31, 2, 00)) == "CET" All of these asserts should pass silently, confirming that the time zone name switches between the right strings at the right times. Observe that even though the return to the standard time zone is scheduled to 3a.m., the moment of the change is marked as 2 a.m. This happens because of a one-hour gap, between 2 a.m. and 3 a.m., that is ambiguous. That hour of time happens twice: once in the time zone CEST, and then again in the time zone CET. Currently, expressing this moment in an unambiguous way, using the standard Python date and time support, is not possible. This is why it is recommended that you store datetime instances in UTC, which is unambiguous, and onlyuse time zone conversion for display purposes. To do the conversion from China to Spain, we've used tzoffset to express the fact that China is eight hours ahead of UTC time (tzoffset is always compared to UTC, not to a particular time zone). Notice how the datetime instance is created with the time zone information. This is always necessary for converting between two different time zones, even if the given time is in the local time zone. If you don't create the instance with the time zone information, you'll get a ValueError: astimezone( ) cannot be applied to a naive datetime. datetime instances are always created naivethey ignore time zone issues entirelyunless you explicitly create them with a time zone. For this purpose, dateutil provides the tzlocal type, which creates instances representing the platform's idea of the local time zone. Besides the types we have seen so far, dateutil also provides tzutc, which creates instances representing UTC; tzfile, which allows using standard binary time zone files; tzical, which creates instances representing iCalendar time zones; and many more besides. See Also Documentation about the dateutil module can be found at https://moin.conectiva.com.br/DateUtil?action=highlight&value=DateUtil, and datetime documentation in the Library Reference. Recipe 3.10. Running a Command Repeatedly Credit: Philip Nunez Problem You need to run a command repeatedly, with arbitrary periodicity. Solution The time.sleep function offers a simple approach to this task: import time, os, sys def main(cmd, inc=60): while True: os.system(cmd) time.sleep(inc) if _ _name_ _ == '_ _main_ _' : numargs = len(sys.argv) - 1 if numargs < 1 or numargs > 2: print "usage: " + sys.argv[0] + " command [seconds_delay]" sys.exit(1) cmd = sys.argv[1] if numargs < 3: main(cmd) else: inc = int(sys.argv[2]) main(cmd, inc) Discussion You can use this recipe with a command that periodically checks for something (e.g., polling), or with one that performs an endlessly repeating action, such as telling a browser to reload a URL whose contents change often, so as to always have a recent version of that URL for viewing. The recipe is structured into a function called main and a body that is preceded by the usual if _ _name_ _=='_ _main_ _': idiom, to execute only if the script runs as a main script. The body examines the command-line arguments you used with the script and calls main appropriately (or gives a usage message if there are too many or too few arguments). This is the best way to structure a script, to make its functionality also available to other scripts that may import it as a module. The main function accepts a cmd string, which is a command you want to pass periodically to the operating system's shell, and, optionally, a period of time in seconds, with a default value of 60 (one minute). main loops forever, alternating between executing the command with os.system and waiting (without consuming resources) with time.sleep. The script's body looks at the command-line arguments you used with the script, which it finds in sys.argv. The first argument, sys.argv[0], is the name of the script, often useful when the script identifies itself as it prints out messages. The body checks that one or two other arguments, in addition to this name, are included. The first (and mandatory) is the command to be run. (You may need to enclose this command in quotes to preserve it from your shell'sparsing: the important thing is that it must remain a single argument.) The second (and optional) argument is the delay in seconds between two runs of the command. If the second argument is missing, the body calls main with just the command argument, accepting the default delay (60 seconds). Note that, if there is a second argument, the body transforms it from a string (all items in sys.argv are always strings) into an integer, which is done most simply by calling built-in type int: inc = int(sys.argv[2]) If the second argument is a string that is not acceptable for transformation into an integer (in other words, if it's anything except a sequence of digits), this call to int raises an exception and terminates the script with appropriate error messages. As one of Python's design principles states, "errors should not pass silently, unless explicitly silenced." It would be bad design to let the script accept an arbitrary string as its second argument, silently taking a default action if that string was not a correct integer representation! For a variant of this recipe that uses the standard Python library module sched, rather than explicit looping and sleeping, see Recipe 3.11. See Also Documentation of the standard library modules os, time, and sys in the Library Reference and Python in a Nutshell; Recipe 3.11. Recipe 3.11. Scheduling Commands Credit: Peter Cogolo Problem You need to schedule commands for execution at certain times. Solution That's what the sched module of the standard library is for: import time, os, sys, sched schedule = sched.scheduler(time.time, time.sleep) def perform_command(cmd, inc): schedule.enter(inc, 0, perform_command, (cmd, inc)) # re-scheduler os.system(cmd) def main(cmd, inc=60): schedule.enter(0, 0, perform_command, (cmd, inc)) # 0==right now schedule.run( ) if _ _name_ _ == '_ _main_ _' : numargs = len(sys.argv) - 1 if numargs < 1 or numargs > 2: print "usage: " + sys.argv[0] + " command [seconds_delay]" sys.exit(1) cmd = sys.argv[1] if numargs < 3: main(cmd) else: inc = int(sys.argv[2]) main(cmd, inc) Discussion This recipe implements the same functionality as in the previous Recipe 3.10, but instead of that recipe's simpler roll-our-own approach, this one uses the standard library module sched. sched is a reasonably simple, yet flexible and powerful, module for scheduling tasks that must take place at given times in the future. To use sched, you first instantiate a scheduler object, such as schedule (shown in this recipe's Solution), with two arguments. The first argument is the function to call in order to find out what time it isnormally time.time, which returns the current time as a number of seconds from an arbitrary reference point known as the epoch. The second argument is the function to call to wait for some timenormally time.sleep. You can also pass functions that measure time in arbitrary artificial ways. For example, you can use sched for such tasks as simulation programs. However, measuring time in artificial ways is an advanced use of sched not covered in this recipe. Once you have a sched.scheduler instance s, you schedule events by calling either s.enter, to schedule something at a relative time n seconds from now (you can pass n as 0 to schedule something for right now), or s.enterabs, to schedule something at a given absolute time. Ineither case, you pass the time (relative or absolute), a priority (if multiple events are scheduled for the same time, they execute in priority order, lowest-priority first), a function to call, and a tuple of arguments to call that function with. Each of these two methods return an event identifier, an arbitrary token that you may store somewhere and later use to cancel a scheduled event by passing the event's token as the argument in a call to s.cancelanother advanced use which this recipe does not cover. After scheduling some events, you call s.run, which keeps running until the queue of scheduled events is empty. In this recipe, we show how to schedule a periodic, recurring event: function perform_command reschedules itself for inc seconds later in the future as the first thing it does, before running the specified system command. In this way, the queue of scheduled events never empties, and function perform_command keeps getting called with regular periodicity. This self- rescheduling is an important idiom, not just in using sched, but any time you have a "one-shot" way to ask for something to happen in the future, and you need instead to have something happen in the future "periodically". (Tkinter's after method, e.g., also works in exactly this way, and thus is also often used with just this kind of self-rescheduling idiom.) Even for a task as simple as the one handled by this recipe, sched still has a small advantage over the simpler roll-your-own approach used previously in Recipe 3.10. In Recipe 3.10, the specified delay occurs between the end of one execution of cmd and the beginning of the next execution. If the execution of cmd takes a highly variable amount of time (as is often the case, e.g., for commands that must wait for the network, or some busy server, etc.), then the command is not really being run periodically. In this recipe, the delay occurs between beginning successive runs of cmd, so that periodicity is indeed guaranteed. If a certain run of cmd takes longer than inc seconds, the schedule temporarily falls behind, but it will eventually catch up again, as long as the average running time of cmd is less than inc seconds: sched never "skips" events. (If you do want an event to be skipped because it's not relevant any more, you have to keep track of the event identifier token and use the cancel method.) For a detailed explanation of this script's structure and body, see Recipe 3.10. See Also Recipe 3.10; documentation of the standard library modules os, time, sys, and sched in the Library Reference and Python in a Nutshell. Recipe 3.12. Doing Decimal Arithmetic Credit: Anna Martelli Ravenscroft Problem You want to perform some simple arithmetic computations in Python 2.4, but you want decimal results, not the Python default of float. Solution To get the normal, expected results from plain, simple computations, use the decimal module introduced in Python 2.4: >>> import decimal >>> d1 = decimal.Decimal('0.3') # assign a decimal-number object >>> d1/3 # try some division Decimal("0.1") >>> (d1/3)*3 # can we get back where we started? Decimal("0.3") Discussion Newcomers to Python (particularly ones without experience with binary float calculations in other programming languages) are often surprised by the results of seemingly simple calculations. For example: >>> f1 = .3 # assign a float >>> f1/3 # try some division 0.099999999999999992 >>> (f1/3)*3 # can we get back where we started? 0.29999999999999999 Binary floating-point arithmetic is the default in Python for very good reasons. You can read all about them in the Python FAQ (Frequently Asked Questions) document at http://www.python.org/doc/faq/general.html#why-are-floating-point-calculations-so-inaccurate, and even in the appendix to the Python Tutorial at http://docs.python.org/tut/node15.html. Many people, however, were unsatisfied with binary floats being the only optionthey wanted to be able to specify the precision, or wanted to use decimal arithmetic for monetary calculations with predictable results. Some of us just wanted the predictable results. (A True Numerical Analyst does, of course, find all results of binary floating-point computations to be perfectly predictable; if any of you three are reading this chapter, you can skip to the next recipe, thanks.) The new decimal type affords a great deal of control over the context for your calculations, allowing you, for example, to set the precision and rounding method to use for the results. However, when all you want is to run simple arithmetical operations that return predictable results, decimal's default context works just fine. Just keep in mind a few points: you may pass a string, integer, tuple, or other decimal object tocreate a new decimal object, but if you have a float n that you want to make into a decimal, pass str(n), not bare n. Also, decimal objects can interact (i.e., be subject to arithmetical operations) with integers, longs, and other decimal objects, but not with floats. These restrictions are anything but arbitrary. Decimal numbers have been added to Python exactly to provide the precision and predictability that float lacks: if it was allowed to build a decimal number from a float, or by operating with one, the whole purpose would be defeated. decimal objects, on the other hand, can be coerced into other numeric types such as float, long, and int, just as you would expect. Keep in mind that decimal is still floating point, not fixed point. If you want fixed point, take a look at Tim Peter's FixedPoint at http://fixedpoint.sourceforge.net/. Also, no money data type is yet available in Python, although you can look at Recipe 3.13 to learn how to roll-your-own money formatting on top of decimal. Last but not least, it is not obvious (at least not to me), when an intermediate computation produces more digits than the inputs, whether you should keep the extra digits for further intermediate computations, and round only when you're done computing a formula (and are about to display or store a result), or whether you should instead round at each step. Different textbooks suggest different answers. I tend to do the former, simply because it's more convenient. If you're stuck with Python 2.3, you may still take advantage of the decimal module, by downloading and installing it as a third-party extensionsee http://www.taniquetil.com.ar/facundo/bdvfiles/get_decimal.html. See Also The explanation of floating-point arithmetic in Appendix B of the Python Tutorial at http://docs.python.org/tut/node15.html; the Python FAQ at http://www.python.org/doc/faq/general.html#why-are-floating-point-calculations-so-inaccurate; Tim Peter's FixedPoint at http://fixedpoint.sourceforge.net/; using decimal as currency, see Recipe 3.13; decimal is documented in the Python 2.4 Library Reference and is available for download to use with 2.3 at http://cvs.sourceforge.net/viewcvs.py/python/python/dist/src/Lib/decimal.py; the decimal PEP (Python Enhancement Proposal), PEP 327, is at http://www.python.org/peps/pep-0327.html. Recipe 3.13. Formatting Decimals as Currency Credit: Anna Martelli Ravenscroft, Alex Martelli, Raymond Hettinger Problem You want to do some tax calculations and display the result in a simple report as Euro currency. Solution Use the new decimal module, along with a modified moneyfmt function (the original, by Raymond Hettinger, is part of the Python library reference section about decimal): import decimal """ calculate Italian invoice taxes given a subtotal. """ def italformat(value, places=2, curr='EUR', sep='.', dp=',', pos='', neg='-', overall=10): """ Convert Decimal ``value'' to a money-formatted string. places: required number of places after the decimal point curr: optional currency symbol before the sign (may be blank) sep: optional grouping separator (comma, period, or blank) every 3 dp: decimal point indicator (comma or period); only specify as blank when places is zero pos: optional sign for positive numbers: "+", space or blank neg: optional sign for negative numbers: "-", "(", space or blank overall: optional overall length of result, adds padding on the left, between the currency symbol and digits """ q = decimal.Decimal((0, (1,), -places)) # 2 places --> '0.01' sign, digits, exp = value.quantize(q).as_tuple( ) result = [ ] digits = map(str, digits) append, next = result.append, digits.pop for i in range(places): if digits: append(next( )) else: append('0') append(dp) i = 0 while digits: append(next( )) i += 1 if i == 3 and digits: i = 0 append(sep) while len(result) < overall: append(' ') append(curr) if sign: append(neg) else: append(pos) result.reverse( ) return ''.join(result) # get the subtotal for use in calculations def getsubtotal(subtin=None): if subtin == None: subtin = input("Enter the subtotal: ") subtotal = decimal.Decimal(str(subtin)) print "\n subtotal: ", italformat(subtotal) return subtotal # specific Italian tax law functions def cnpcalc(subtotal): contrib = subtotal * decimal.Decimal('.02') print "+ contributo integrativo 2%: ", italformat(contrib, curr='') return contrib def vatcalc(subtotal, cnp): vat = (subtotal+cnp) * decimal.Decimal('.20') print "+ IVA 20%: ", italformat(vat, curr='') return vat def ritacalc(subtotal): rit = subtotal * decimal.Decimal('.20') print "-Ritenuta d'acconto 20%: ", italformat(rit, curr='') return rit def dototal(subtotal, cnp, iva=0, rit=0): totl = (subtotal+cnp+iva)-rit print " TOTALE: ", italformat(totl) return totl # overall calculations report def invoicer(subtotal=None, context=None): if context is None: decimal.getcontext( ).rounding="ROUND_HALF_UP" # Euro rounding rules else: decimal.setcontext(context) # set to context arg subtot = getsubtotal(subtotal) contrib = cnpcalc(subtot) dototal(subtot, contrib, vatcalc(subtot, contrib), ritacalc(subtot)) if _ _name_ _=='_ _main_ _': print "Welcome to the invoice calculator" tests = [100, 1000.00, "10000", 555.55] print "Euro context" for test in tests: invoicer(test) print "default context" for test in tests: invoicer(test, context=decimal.DefaultContext) Discussion Italian tax calculations are somewhat complicated, more so than this recipe demonstrates. This recipe applies only to invoicing customers within Italy. I soon got tired of doing them by hand, so I wrote a simple Python script to do the calculations for me. I've currently refactored into the version shown in this recipe, using the new decimal module, just on the principle that money computations should never, but never, be done with binary floats. How to best use the new decimal module for monetary calculations was not immediately obvious. While the decimal arithmetic is pretty straightforward, the options for displaying results were less clear. The italformat function in the recipe is based on Raymond Hettinger's moneyfmt recipe, found in the decimal module documentation available in the Python 2.4 Library Reference. Some minor modifications were helpful for my reporting purposes. The primary addition was the overall parameter. This parameter builds a decimal with a specific number ofoverall digits, with whitespace padding between the currency symbol (if any) and the digits. This eases alignment issues when the results are of a standard, predictable length. Notice that I have coerced the subtotal input subtin to be a string in subtotal = decimal.Decimal(str(subtin)). This makes it possible to feed floats (as well as integers or strings) to getsubtotal without worrywithout this, a float would raise an exception. If your program is likely to pass tuples, refactor the code to handle that. In my case, a float was a rather likely input to getsubtotal, but I didn't have to worry about tuples. Of course, if you need to display using U.S. $, or need to use other rounding rules, it's easy enough to modify things to suit your needs. For example, to display U.S. currency, you could change the curr, sep, and dp arguments' default values as follows: def USformat(value, places=2, curr='$', sep=',', dp='.', pos='', neg='-', overall=10): ... If you regularly have to use multiple currency formats, you may choose to refactor the function so that it looks up the appropriate arguments in a dictionary, or you may want to find other ways to pass the appropriate arguments. In theory, the locale module in the Python Standard Library should be the standard way to let your code access locale-related preferences such as those connected to money formatting, but in practice I've never had much luck using locale (for this or any other purpose), so that's one task that I'll gladly leave as an exercise to the reader. Countries often have specific rules on rounding; decimal uses ROUND_HALF_EVEN as the default. However, the Euro rules specify ROUND_HALF_UP. To use different rounding rules, change the context, as shown in the recipe. The result of this change may or may not be obvious, but one should be aware that it can make a (small, but legally not negligible) difference. You can also change the context more extensively, by creating and setting your own context class instance. A change in context, whether set by a simple getcontext attribution change, or with a custom context class instance passed to setcontext(mycontext), continues to apply throughout the active thread, until you change it. If you are considering using decimal in production code (or even for your own home bookkeeping use), be sure to use the right context (in particular, the correct rounding rules) for your country's accounting practices. See Also Python 2.4's Library Reference on decimal, particularly the section on decimal.context and the "recipes" at the end of that section. Recipe 3.14. Using Python as a Simple Adding Machine Credit: Brett Cannon Problem You want to use Python as a simple adding machine, with accurate decimal (not binary floating- point!) computations and a "tape" that shows the numbers in an uncluttered columnar view. Solution To perform the computations, we can rely on the decimal module. We accept input lines, each made up of a number followed by an arithmetic operator, an empty line to request the current total, and q to terminate the program: import decimal, re, operator parse_input = re.compile(r'''(?x) # allow comments and whitespace in the RE (\d+\.?\d*) # number with optional decimal part \s* # optional whitespace ([-+/*]) # operator $''') # end-of-string oper = { '+': operator.add, '-': operator.sub, '*': operator.mul, '/': operator.truediv, } total = decimal.Decimal('0') def print_total( ): print '== == =\n', total print """Welcome to Adding Machine: Enter a number and operator, an empty line to see the current subtotal, or q to quit: """ while True: try: tape_line = raw_input( ).strip( ) except EOFError: tape_line = 'q' if not tape_line: print_total( ) continue elif tape_line == 'q': print_total( ) break try: num_text, op = parse_input.match(tape_line).groups( ) except AttributeError: print 'Invalid entry: %r' % tape_line print 'Enter number and operator, empty line for total, q to quit' continue total = oper[op](total, decimal.Decimal(num_text)) DiscussionPython's interactive interpreter is often a useful calculator, but a simpler "adding machine" also has its uses. For example, an expression such as 2345634+2894756-2345823 is not easy to read, so checking that you're entering the right numbers for a computation is not all that simple. An adding machine's tape shows numbers in a simple, uncluttered columnar view, making it easier to double check what you have entered. Moreover, the decimal module performs computations in the normal, decimal-based way we need in real life, rather than in the floating- point arithmetic preferred by scientists, engineers, and today's computers. When you run the script in this recipe from a normal command shell (this script is not meant to be run from within a Python interactive interpreter!), the script prompts you once, and then just sits there, waiting for input. Type a number (one or more digits, then optionally a decimal point, then optionally more digits), followed by an operator (/, *, -, or + the four operator characters you find on the numeric keypad on your keyboard), and then press return. The script applies the number to the running total using the operator. To output the current total, just enter a blank line. To quit, enter the letter q and press return. This simple interface matches the input/output conventions of a typical simple adding machine, removing the need to have some other form of output. The decimal package is part of Python's standard library since version 2.4. If you're still using Python 2.3, visit http://www.taniquetil.com.ar/facundo/bdvfiles/get_decimal.html and download and install the package in whatever form is most convenient for you. decimal allows high- precision decimal arithmetic, which is more convenient for many uses (such as any computation involving money) than the binary floating-point computations that are faster on today's computers and which Python uses by default. No more lost pennies due to hard-to-understand issues with binary floating point! As demonstrated in Recipe 3.13, you can even change the rounding rules from the default of ROUND_HALF_EVEN, if you really need to. This recipe's script is meant to be very simple, so many improvements are possible. A useful enhancement would be to keep the "tape" on disk for later checking. You can do that easily, by adding, just before the loop, a statement to open some appropriate text file for append: tapefile = open('tapefile.txt', 'a') and, just after the try/except statement that obtains a value for tape_line, a statement to write that value to the file: tapefile.write(tape_line+'\n') If you do want to make these additions, you will probably also want to enrich function print_total so that it writes to the "tape" file as well as to the command window, therefore, change the function to: def print_total( ): print '== == =\n', total tapefile.write('== == =\n' + str(total) + '\n') The write method of a file object accepts a string as its argument and does not implicitly terminate the line as the print statement does, so we need to explicitly call the str built-in function and explicitly add '\n' as needed. Alternatively, the second statement in this version of print_total could be coded in a way closer to the first one: print >>tapefile, '== == =\n', total Some people really dislike this print >>somefile, syntax, but it can come in handy in cases such as this one.More ambitious improvements would be to remove the need to press Return after each operator (that would require performing unbuffered input and dealing with one character at a time, rather than using the handy but line-oriented built-in function raw_input as the recipe doessee Recipe 2.23 for a cross-platform way to get unbuffered input), to add a clear function (or clarify to users that inputting 0* will zero out the "tape"), and even to add a GUI that looks like an adding machine. However, I'm leaving any such improvements as exercises for the reader. One important point about the recipe's implementation is the oper dictionary, which uses operator characters (/, *, -, +) as keys and the appropriate arithmetic functions from the built-in module operator, as corresponding values. The same effect could be obtained, more verbosely, by a "tree" of if/elif, such as: if op == '+': total = total + decimal.Decimal(num_text) elif op == '-': total = total - decimal.Decimal(num_text) elif op == '*': ... and so on ... However, Python dictionaries are very idiomatic and handy for such uses, and they lead to less repetitious and thus more maintainable code. See Also decimal is documented in the Python 2.4 Library Reference, and is available for download to use with 2.3 at http://www.taniquetil.com.ar/facundo/bdvfiles/get_decimal.html; you can read the decimal PEP 327 at http://www.python.org/peps/pep-0327.html. Recipe 3.15. Checking a Credit Card Checksum Credit: David Shaw, Miika Keskinen Problem You need to check whether a credit card number respects the industry standard Luhn checksum algorithm. Solution Luhn mod 10 is the credit card industry's standard for credit card checksums. It's not built into Python, but it's easy to roll our own computation for it: def cardLuhnChecksumIsValid(card_number): """ checks to make sure that the card passes a luhn mod-10 checksum """ sum = 0 num_digits = len(card_number) oddeven = num_digits & 1 for count in range(num_digits): digit = int(card_number[count]) if not (( count & 1 ) ^ oddeven): digit = digit * 2 if digit > 9: digit = digit - 9 sum = sum + digit return (sum % 10) == 0 Discussion This recipe was originally written for a now-defunct e-commerce application to be used within Zope. It can save you time and money to apply this simple validation before trying to process a bad or miskeyed card with your credit card vendor, because you won't waste money trying to authorize a bad card number. The recipe has wider applicability because many government identification numbers also use the Luhn (i.e., modulus 10) algorithm. A full suite of credit card validation methods is available at http://david.theresistance.net/files/creditValidation.py If you're into cool one-liners rather than simplicity and clarity, (a) you're reading the wrong book (the Perl Cookbook is a great book that will make you much happier), (b) meanwhile, to keep you smiling while you go purchase a more appropriate oeuvre, try: checksum = lambda a: ( 10 - sum([int(y)*[7,3,1][x%3] for x, y in enumerate(str(a)[::-1])])%10)%10 See AlsoA good therapist, if you do prefer the one-line checksum version. Recipe 3.16. Watching Foreign Exchange Rates Credit: Victor Yongwei Yang Problem You want to monitor periodically (with a Python script to be run by crontab or as a Windows scheduled task) an exchange rate between two currencies, obtained from the Web, and receive email alerts when the rate crosses a certain threshold. Solution This task is similar to other monitoring tasks that you could perform on numbers easily obtained from the Web, be they exchange rates, stock quotes, wind-chill factors, or whatever. Let's see specifically how to monitor the exchange rate between U.S. and Canadian dollars, as reported by the Bank of Canada web site (as a simple CSV (comma-separated values) feed that is easy to parse): import httplib import smtplib # configure script's parameters here thresholdRate = 1.30 smtpServer = 'smtp.freebie.com' fromaddr = 'foo@bar.com' toaddrs = 'your@corp.com' # end of configuration url = '/en/financial_markets/csv/exchange_eng.csv' conn = httplib.HTTPConnection('www.bankofcanada.ca') conn.request('GET', url) response = conn.getresponse( ) data = response.read( ) start = data.index('United States Dollar') line = data[start:data.index('\n', start)] # get the relevant line rate = line.split(',')[-1] # last field on the line if float(rate) < thresholdRate: # send email msg = 'Subject: Bank of Canada exchange rate alert %s' % rate server = smtplib.SMTP(smtpServer) server.sendmail(fromaddr, toaddrs, msg) server.quit( ) conn.close( ) Discussion When working with foreign currencies, it is particularly useful to have an automated way of getting the conversions you need. This recipe provides this functionality in a quite simple, straightforward manner. When cron runs this script, the script goes to the site, and gets the CSV feed, which provides the daily noon exchange rates for the previous seven days: Date (m/d/year),11/12/2004,11/15/2004, ... ,11/19/2004,11/22/2004$Can/US closing rate,1.1927,1.2005,1.1956,1.1934,1.2058,1.1930, United States Dollar,1.1925,1.2031,1.1934,1.1924,1.2074,1.1916,1.1844 ... The script then continues to find the specific currency ('United States Dollar') and reads the last field to find today's rate. If you're having trouble understanding how that works, it may be helpful to break it down: US = data.find('United States Dollar') # find the index of the currency endofUSline = data.index('\n', US) # find index for that line end USline = data[US:endofUSline] # slice to make one string rate = USline.split(',')[-1] # split on ',' and return last field The recipe provides an email alert when the rate falls below a particular threshold, which can be configured to whatever rate you prefer (e.g., you could change that statement to send you an alert whenever the rate changes outside a threshold range). See Also httplib, smtplib, and string function are documented in the Library Reference and Python in a Nutshell. Chapter 4. Python Shortcuts Introduction Recipe 4.1. Copying an Object Recipe 4.2. Constructing Lists with List Comprehensions Recipe 4.3. Returning an Element of a List If It Exists Recipe 4.4. Looping over Items and Their Indices in a Sequence Recipe 4.5. Creating Lists of Lists Without Sharing References Recipe 4.6. Flattening a Nested Sequence Recipe 4.7. Removing or Reordering Columnsin a List of Rows Recipe 4.8. Transposing Two-Dimensional Arrays Recipe 4.9. Getting a Value from a Dictionary Recipe 4.10. Adding an Entry to a Dictionary Recipe 4.11. Building a Dictionary Without Excessive Quoting Recipe 4.12. Building a Dict from a List of Alternating Keys and Values Recipe 4.13. Extracting a Subset of a Dictionary Recipe 4.14. Inverting a Dictionary Recipe 4.15. Associating Multiple Values with Each Key in a Dictionary Recipe 4.16. Using a Dictionary to Dispatch Methods or Functions Recipe 4.17. Finding Unions and Intersections of Dictionaries Recipe 4.18. Collecting a Bunch of Named Items Recipe 4.19. Assigning and Testing with One Statement Recipe 4.20. Using printf in Python Recipe 4.21. Randomly Picking Items with Given Probabilities Recipe 4.22. Handling Exceptions Within an Expression Recipe 4.23. Ensuring a Name Is Defined in a Given Module Introduction Credit: David Ascher, ActiveState, co-author of Learning Python Programming languages are like natural languages. Each has a set of qualities that polyglots generally agree on as characteristics of the language. Russian and French are often admired for their lyricism, while English is more often cited for its precision and dynamism: unlike the Académie-defined French language, the English language routinely grows words to suit its speakers' needs, such as "carjacking," "earwitness," "snailmail," "email," "googlewhacking," and "blogging." In the world of computer languages, Perl is well known for its many degrees of freedom: TMTOWTDI (There's More Than One Way To Do It) is one of the mantras of the Perl programmer. Conciseness is also seen as a strong virtue in the Perl and APL communities. As you'll see in many of the discussions of recipes throughout this volume, in contrast, Python programmers often express their belief in the value of clarity and elegance. As a well-known Perl hacker once told me, Python's prettier, but Perl is more fun. I agree with him that Python does have a strong (as in well-defined) aesthetic, while Perl has more of a sense of humor. The reason I mention these seemingly irrelevant characteristics at the beginning of this chapter is that the recipes you see in this chapter are directly related to Python's aesthetic and social dynamics. If this book had been about Perl, the recipes in a shortcuts chapter would probably elicit head scratching, contemplation, an "a-ha"! moment, and then a burst of laughter, as the reader grokked the genius behind a particular trick. In contrast, in most of the recipes in this chapter, the author presents a single elegant language feature, but one that he feels is underappreciated. Much like I, a proud resident of Vancouver, will go out of my way to show tourists the really neat things about the city, from the parks to the beaches to the mountains, a Python user will seek out friends and colleagues and say, "You gotta see this!" For me and most of the programmers I know, programming in Python is a shared social pleasure, not a competitive pursuit. There is great pleasure in learning a new feature and appreciating its design, elegance, and judicious use, and there's a twin pleasure in teaching another or another thousand about that feature. A word about the history of the chapter: back when we identified the recipe categories for the first edition of this collection, our driving notion was that there would be recipes of various kinds, each with a specific goala soufflé, a tart, an osso buco. Those recipes would naturally fall into fairly typical categories, such as desserts, appetizers, and meat dishes, or their perhaps less appetizing, nonmetaphorical equivalents, such as files, algorithms, and so on. So we picked a list of categories, added the categories to the Zope site used to collect recipes, and opened the floodgates. Soon, it became clear that some submissions were hard to fit into the predetermined categories. There's a reason for that, and cooking helps explain why. The recipes in this chapter are the Pythonic equivalent of making a roux (a cooked mixture of fat and flour, used in making sauces, for those of you without a classic French cooking background), kneading dough, flouring, separating eggs, flipping a pan's contents, blanching, and the myriad other tricks that any accomplished cook knows, but that you won't find in a typical cookbook. Many of these tricks and techniques are used in preparing meals, but it's hard to pigeonhole them as relevant for a given type of dish. And if you're a novice cook looking up a fancy recipe, you're likely to get frustrated quickly because serious cookbook authors assume you know these techniques, and they explain them (with illustrations!) only in books with titles such as Cooking for Divorced Middle-Aged Men. We didn't want to exclude this precious category of tricks from this book, so a new category was born (sorry, no illustrations). In the introduction to this chapter in the first edition, I presciently said: I believe that the recipes in this chapter are among the most time-sensitive of the recipes in this volume. That's because the aspects of the language that people consider shortcuts ornoteworthy techniques seem to be relatively straightforward, idiomatic applications of recent language features. I can proudly say that I was right. This new edition, significantly focused on the present definition of the language, makes many of the original recipes irrelevant. In the two Python releases since the book's first edition, Python 2.3 and 2.4, the language has evolved to incorporate the ideas of those recipes into new syntactic features or library functions, just as it had done with every previous major release, making a cleaner, more compact, and yet more powerful language that's as much fun to use today as it was over ten years ago. All in all, about half the recipes in this chapter (roughly the same proportion as in the rest of the book) are entirely new ones, while the other half are vastly revised (mostly simplified) versions of recipes that were in the first edition. Thanks to the simplifications, and to the focus on just two language versions (2.3 and 2.4) rather than the whole panoply of older versions that was covered by the first edition, this chapter, as well as the book as a whole, has over one-third more recipes than the first edition did. It's worth noting in closing that many of the recipes that are in this newly revised chapter touch on some of the most fundamental, unchanging aspects of the language: the semantics of assignment, binding, copy, and references; sequences; dictionaries. These ideas are all keys to the Pythonic approach to programming, and seeing these recipes live for several years makes me wonder whether Python will evolve in the next few years in related directions. Recipe 4.1. Copying an Object Credit: Anna Martelli Ravenscroft, Peter Cogolo Problem You want to copy an object. However, when you assign an object, pass it as an argument, or return it as a result, Python uses a reference to the original object, without making a copy. Solution Module copy in the standard Python library offers two functions to create copies. The one you should generally use is the function named copy, which returns a new object containing exactly the same items and attributes as the object you're copying: import copy new_list = copy.copy(existing_list) On the rare occasions when you also want every item and attribute in the object to be separately copied, recursively, use deepcopy: import copy new_list_of_dicts = copy.deepcopy(existing_list_of_dicts) Discussion When you assign an object (or pass it as an argument, or return it as a result), Python (like Java) uses a reference to the original object, not a copy. Some other programming languages make copies every time you assign something. Python never makes copies "implicitly" just because you're assigning: to get a copy, you must specifically request a copy. Python's behavior is simple, fast, and uniform. However, if you do need a copy and do not ask for one, you may have problems. For example: >>> a = [1, 2, 3] >>> b = a >>> b.append(5) >>> print a, b [1, 2, 3, 5] [1, 2, 3, 5] Here, the names a and b both refer to the same object (a list), so once we alter the object through one of these names, we later see the altered object no matter which name we use for it. No original, unaltered copy is left lying about anywhere.To become an effective Python programmer, it is crucial that you learn to draw the distinction between altering an object and assigning to a name, which previously happened to refer to the object. These two kinds of operations have nothing to do with each other. A statement such as a=[ ] rebinds name a but performs no alteration at all on the object that was previously bound to name a. Therefore, the issue of references versus copies just doesn't arise in this case: the issue is meaningful only when you alter some object. If you are about to alter an object, but you want to keep the original object unaltered, you must make a copy. As this recipe's solution explains, the module copy from the Python Standard Library offers two functions to make copies. Normally, you use copy.copy, which makes a shallow copyit copies an object, but for each attribute or item of the object, it continues to share references, which is faster and saves memory. Shallow copying, alas, isn't sufficient to entirely "decouple" a copied object from the original one, if you propose to alter the items or attributes of either object, not just the object itself: >>> list_of_lists = [ ['a'], [1, 2], ['z', 23] ] >>> copy_lol = copy.copy(lists_of_lists) >>> copy_lol[1].append('boo') >>> print list_of_lists, copy_lol [['a'], [1, 2, 'boo'], ['z', 23]] [['a'], [1, 2, 'boo'], ['z', 23]] Here, the names list_of_lists and copy_lol refer to distinct objects (two lists), so we could alter either of them without affecting the other. However, each item of list_of_lists is the same object as the corresponding item of copy_lol, so once we alter an item reached by indexing either of these names, we later see the altered item no matter which object we're indexing to reach it. If you do need to copy some container object and also recursively copy all objects it refers to (meaning all items, all attributes, and also items of items, items of attributes, etc.), use copy.deepcopysuch deep copying may cost you substantial amounts of time and memory, but if you gotta, you gotta. For deep copies, copy.deepcopy is the only way to go. For normal shallow copies, you may have good alternatives to copy.copy, if you know the type of the object you want to copy. To copy a list L, call list(L); to copy a dict d, call dict(d); to copy a set s (in Python 2.4, which introduces the built-in type set), call set(s). (Since list, dict, and, in 2.4, set, are built-in names, you do not need to perform any "preparation" before you use any of them.) You get the general pattern: to copy a copyable object o, which belongs to some built-in Python type t, you may generally just call t(o). dicts also offer a dedicated method to perform a shallow copy: d.copy( ) and dict(d) do the same thing. Of the two, I suggest you use dict(d): it's more uniform with respect to other types, and it's even shorter by one character! To copy instances of arbitrary types or classes, whether you coded them or got them from a library, just use copy.copy. If you code your own classes, it's generally not worth the bother to define your own copy or clone method. If you want to customize the way instances of your class get (shallowly) copied, your class can supply a special method _ _copy_ _ (see Recipe 6.9 for a special technique relating to the implementation of such a method), or special methods _ _getstate_ _ and _ _setstate_ _. (See Recipe 7.4 for notes on these special methods, which also help with deep copying and serializationi.e., picklingof instances of your class.) If you want to customize the way instances of your class get deeply copied, your class can supply a special method _ _deepcopy_ _ (see Recipe 6.9.) Note that you do not need to copy immutable objects (strings, numbers, tuples, etc.) because you don't have to worry about altering them. If you do try to perform such a copy, you'll just getthe original right back; no harm done, but it's a waste of time and code. For example: >>> s = 'cat' >>> t = copy.copy(s) >>> s is t True The is operator checks whether two objects are not merely equal, but in fact the same object (is checks for identity; for checking mere equality, you use the == operator). Checking object identity is not particularly useful for immutable objects (we're using it here just to show that the call to copy.copy was useless, although innocuous). However, checking object identity can sometimes be quite important for mutable objects. For example, if you're not sure whether two names a and b refer to separate objects, or whether both refer to the same object, a simple and very fast check a is b lets you know how things stand. That way you know whether you need to copy the object before altering it, in case you want to keep the original object unaltered. You can use other, inferior ways exist to create copies, namely building your own. Given a list L, both a "whole-object slice" L[:] and a list comprehension [x for x in L] do happen to make a (shallow) copy of L, as do adding an empty list, L+[ ], and multiplying the list by 1, L*1 . . . but each of these constructs is just wasted effort and obfuscationcalling list(L) is clearer and faster. You should, however, be familiar with the L[:] construct because for historical reasons it's widely used. So, even though you're best advised not to use it yourself, you'll see it in Python code written by others. Similarly, given a dictionary d, you could create a shallow copy named d1 by coding out a loop: >>> d1 = { } >>> for somekey in d: ... d1[somekey] = d[somekey] or more concisely by d1 = { }; d1.update(d). However, again, such coding is a waste of time and effort and produces nothing but obfuscated, fatter, and slower code. Use d1=dict(d), be happy! See Also Module copy in the Library Reference and Python in a Nutshell. Recipe 4.2. Constructing Lists with List Comprehensions Credit: Luther Blissett Problem You want to construct a new list by operating on elements of an existing sequence (or other kind of iterable). Solution Say you want to create a new list by adding 23 to each item of some other list. A list comprehension expresses this idea directly: thenewlist = [x + 23 for x in theoldlist] Similarly, say you want the new list to comprise all items in the other list that are larger than 5. A list comprehension says exactly that: thenewlist = [x for x in theoldlist if x > 5] When you want to combine both ideas, you can perform selection with an if clause, and also use some expression, such as adding 23, on the selected items, in a single pass: thenewlist = [x + 23 for x in theoldlist if x > 5] Discussion Elegance, clarity, and pragmatism, are Python's core values. List comprehensions show how pragmatism can enhance both clarity and elegance. Indeed, list comprehensions are often the best approach even when, instinctively, you're thinking not of constructing a new list but rather of "altering an existing list". For example, if your task is to set all items greater than 100 to 100, in an existing list object L, the best solution is: L[:] = [min(x,100) for x in L] Assigning to the "whole-list slice" L[:] alters the existing list object in place, rather than just rebinding the name L, as would be the case if you coded L = . . . instead. You should not use a list comprehension when you simply want to perform a loop. When you want a loop, code a loop. For an example of looping over a list, see Recipe 4.4. See Chapter 19 for more information about iteration in Python. It's also best not to use a list comprehension when another built-in does what you want even more directly and immediately. For example, to copy a list, use L1 = list(L), not:L1 = [x for x in L] Similarly, when the operation you want to perform on each item is to call a function on the item and use the function's result, use L1 = map(f, L) rather than L1 = [f(x) for x in L]. But in most cases, a list comprehension is just right. In Python 2.4, you should consider using a generator expression, rather than a list comprehension, when the sequence may be long and you only need one item at a time. The syntax of generator expressions is just the same as for list comprehensions, except that generator expressions are surrounded by parentheses, ( and ), not brackets, [ and ]. For example, say that we only need the summation of the list computed in this recipe's Solution, not each item of the list. In Python 2.3, we would code: total = sum([x + 23 for x in theoldlist if x > 5]) In Python 2.4, we can code more naturally, omitting the brackets (no need to add additional parenthesesthe parentheses already needed to call the built-in sum suffice): total = sum(x + 23 for x in theoldlist if x > 5) Besides being a little bit cleaner, this method avoids materializing the list as a whole in memory and thus may be slightly faster when the list is extremely long. See Also The Reference Manual section on list displays (another name for list comprehensions) and Python 2.4 generator expressions; Chapter 19; the Library Reference and Python in a Nutshell docs on the itertools module and on the built-in functions map, filter, and sum; Haskell is at http://www.haskell.org. Python borrowed list comprehensions from the functional language Haskell (http://www.haskell.org), changing the syntax to use keywords rather than punctuation. If you do know Haskell, though, take care! Haskell's list comprehensions, like the rest of Haskell, use lazy evaluation (also known as normal order or call by need). Each item is computed only when it's needed. Python, like most other languages, uses (for list comprehensions as well as elsewhere) eager evaluation (also known as applicative order, call by value, or strict evaluation). That is, the entire list is computed when the list comprehension executes, and kept in memory afterwards as long as necessary. If you are translating into Python a Haskell program that uses list comprehensions to represent infinite sequences, or even just long sequences of which only one item at a time must be kept around, Python list comprehensions may not be suitable. Rather, look into Python 2.4's new generator expressions, whose semantics are closer to the spirit of Haskell's lazy evaluationeach item gets computed only when needed. Recipe 4.3. Returning an Element of a List If It Exists Credit: Nestor Nissen, A. Bass Problem You have a list L and an index i, and you want to get L[i] when i is a valid index into L; otherwise, you want to get a default value v. If L were a dictionary, you'd use L.get(i, v), but lists don't have a get method. Solution Clearly, we need to code a function, and, in this case, the simplest and most direct approach is the best one: def list_get(L, i, v=None): if -len(L) <= i < len(L): return L[i] else: return v Discussion The function in this recipe just checks whether i is a valid index by applying Python's indexing rule: valid indices are negative ones down to -len(L) inclusive, and non-negative ones up to len(L) exclusive. If almost all calls to list_get pass a valid index value for i, you might prefer an alternative approach: def list_get_egfp(L, i, v=None): try: return L[i] except IndexError: return v However, unless a vast majority of the calls pass a valid index, this alternative (as some time- measurements show) can be up to four times slower than the list_get function shown in the solution. Therefore, this "easier to get forgiveness than permission" (EGFP) approach, although it is often preferable in Python, cannot be recommended for this specific case. I've also tried quite a few fancy, intricate and obscure approaches, but, besides being hard to explain and to understand, they all end up slower than the plain, simple function list_get. General principle: when you write Python code, prefer clarity and readability to compactness and tersenesschoose simplicity over subtlety. You often will be rewarded with code that runs faster, and invariably, you will end up with code that is less prone to bugs and is easier to maintain, which is far more important than minor speed differences in 99.9% of the cases you encounter in the real world. See Also Language Reference and Python in a Nutshell documentation on list indexing. Recipe 4.4. Looping over Items and Their Indices in a Sequence Credit: Alex Martelli, Sami Hangaslammi Problem You need to loop on a sequence, but at each step you also need to know which index into the sequence you have reached (e.g., because you need to rebind some entries in the sequence), and Python's preferred approach to looping doesn't use the indices. Solution That's what built-in function enumerate is for. For example: for index, item in enumerate(sequence): if item > 23: sequence[index] = transform(item) This is cleaner, more readable, and faster than the alternative of looping over indices and accessing items by indexing: for index in range(len(sequence)): if sequence[index] > 23: sequence[index] = transform(sequence[index]) Discussion Looping on a sequence is a very frequent need, and Python strongly encourages you to do just that, looping on the sequence directly. In other words, the Pythonic way to get each item in a sequence is to use: for item in sequence: process(item) rather than the indirect approach, typical of lower-level languages, of looping over the sequence's indices and using each index to fetch the corresponding item: for index in range(len(sequence)): process(sequence[index]) Looping directly is cleaner, more readable, faster, and more general (since you can loop on any iterable, by definition, while indexing works only on sequences, such as lists). However, sometimes you do need to know the index, as well as the corresponding item, within the loop. The most frequent reason for this need is that, in order to rebind an entry in a list, you must assign the new item to thelist[index]. To support this need, Python offers the built-infunction enumerate, which takes any iterable argument and returns an iterator yielding all the pairs (two-item tuples) of the form (index, item), one pair at a time. By writing your for loop's header clause in the form: for index, item in enumerate(sequence): both the index and the item are available within the loop's body. For help remembering the order of the items in each pair enumerate yields, think of the idiom d=dict(enumerate(L)). This gives a dictionary d that's equivalent to list L, in the sense that d[i] is L[i] for any valid non-negative index i. See Also Library Reference and Python in a Nutshell section about enumerate; Chapter 19. Recipe 4.5. Creating Lists of Lists Without Sharing References Credit: David Ascher Problem You want to create a multidimensional list but want to avoid implicit reference sharing. Solution To build a list and avoid implicit reference sharing, use a list comprehension. For example, to build a 5 x 10 array of zeros: multilist = [[0 for col in range(5)] for row in range(10)] Discussion When a newcomer to Python is first shown that multiplying a list by an integer repeats that list that many times, the newcomer often gets quite excited about it, since it is such an elegant notation. For example: >>> alist = [0] * 5 is clearly an excellent way to get an array of 5 zeros. The problem is that one-dimensional tasks often grow a second dimension, so there is a natural progression to: >>> multi = [[0] * 5] * 3 >>> print multi [[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]] This appears to work, but the same newcomer is then often puzzled by bugs, which typically can be boiled down to a snippet such as: >>> multi[0][0] = 'oops!' >>> print multi [['oops!', 0, 0, 0, 0], ['oops!', 0, 0, 0, 0], ['oops!', 0, 0, 0, 0]] This issue confuses most programmers at least once, if not a few times (see the FAQ entry at http://www.python.org/doc/FAQ.html#4.50). To understand the issue, it helps to decompose the creation of the multidimensional list into two steps: >>> row = [0] * 5 # a list with five references to 0 >>> multi = [row] * 3 # a list with three references to the row objectThis decomposed snippet produces a multi that's identical to that given by the more concise snippet [[0]*5]*3 shown earlier, and it has exactly the same problem: if you now assign a value to multi[0][0], you have also changed the value of multi[1][0] and that of multi[2][0] . . . , and, indeed, you have changed the value of row[0], too! The comments are key to understanding the source of the confusion. Multiplying a sequence by a number creates a new sequence with the specified number of new references to the original contents. In the case of the creation of row, it doesn't matter whether or not references are being duplicated, since the referent (the object being referred to) is a number, and therefore immutable. In other words, there is no practical difference between an object and a reference to an object if that object is immutable. In the second line, however, we create a new list containing three references to the contents of the [row] list, which holds a single reference to a list. Thus, multi contains three references to a single list object. So, when the first element of the first element of multi is changed, you are actually modifying the first element of the shared list. Hence the surprise. List comprehensions, as shown in the "Solution", avoid the problem. With list comprehensions, no sharing of references occursyou have a truly nested computation. If you have followed the discussion thoroughly, it may have occurred to you that we don't really need the inner list comprehension, only the outer one. In other words, couldn't we get just the same effect with: multilist = [[0]*5 for row in range(10)] The answer is that, yes, we could, and in fact using list multiplication for the innermost axis and list comprehension for all outer ones is fasterover twice as fast in this example. So why don't I recommend this latest solution? Answer: the speed improvement for this example is from 57 down to 24 microseconds in Python 2.3, from 49 to 21 in Python 2.4, on a typical PC of several years ago (AMD Athlon 1.2 GHz CPU, running Linux). Shaving a few tens of microseconds from a list-creation operation makes no real difference to your application's performance: and you should optimize your code, if at all, only where it matters, where it makes a substantial and important difference to the performance of your application as a whole. Therefore, I prefer the code shown in the recipe's Solution, simply because using the same construct for both the inner and the outer list creations makes it more conceptually symmetrical and easier to read! See Also Documentation for the range built-in function in the Library Reference and Python in a Nutshell. Recipe 4.6. Flattening a Nested Sequence Credit: Luther Blissett, Holger Krekel, Hemanth Sethuram, ParzAspen Aspen Problem Some of the items in a sequence may in turn be sub-sequences, and so on, to arbitrary depth of "nesting". You need to loop over a "flattened" sequence, "expanding" each sub-sequence into a single, flat sequence of scalar items. (A scalar, or atom, is anything that is not a sequencei.e., a leaf, if you think of the nested sequence as a tree.) Solution We need to be able to tell which of the elements we're handling are "subsequences" to be "expanded" and which are "scalars" to be yielded as is. For generality, we can take an argument that's a predicate to tell us what items we are to expand. (A predicate is a function that we can call on any element and that returns a truth value: in this case, true if the element is a subsequence we are to expand, False otherwise.) By default, we can arbitrarily say that every list or tuple is to be "expanded", and nothing else. Then, a recursive generator offers the simplest solution: def list_or_tuple(x): return isinstance(x, (list, tuple)) def flatten(sequence, to_expand=list_or_tuple): for item in sequence: if to_expand(item): for subitem in flatten(item, to_expand): yield subitem else: yield item Discussion Flattening a nested sequence, or, equivalently, "walking" sequentially over all the leaves of a "tree", is a common task in many kinds of applications. You start with a nested structure, with items grouped into sequences and subsequences, and, for some purposes, you don't care about the structure at all. You just want to deal with the items, one after the other. For example, for x in flatten([1, 2, [3, [ ], 4, [5, 6], 7, [8,], ], 9]): print x, emits 1 2 3 4 5 6 7 8 9. The only problem with this common task is that, in the general case, determining what is to be "expanded", and what is to be yielded as a scalar, is not as obvious as it might seem. So, I ducked that decision, delegating it to a callable predicate argument that the caller can pass to flatten, unless the caller accepts flatten's somewhat simplistic default behavior of expanding just tuples and lists. In the same module as flatten, we should also supply another predicate that a caller might wellwant to usea predicate that will expand just about any iterable except strings (plain and Unicode). Strings are iterable, but almost invariably applications want to treat them as scalars, not as subsequences. To identify whether an object is iterable, we just need to try calling the built-in iter on that object: the call raises TypeError if the object is not iterable. To identify whether an object is string-like, we simply check whether the object is an instance of basestring, since isinstance(obj, basestring) is true when obj is an instance of any subclass of basestringthat is, any string-like type. So, the alternative predicate is not hard to code: def nonstring_iterable(obj): try: iter(obj) except TypeError: return False else: return not isinstance(obj, basestring) Now the caller may choose to call flatten(seq, nonstring_iterable) when the need is to expand any iterable that is not a string. It is surely better not to make the nonstring_iterable predicate the default for flatten, though: in a simple case, such as the example snippet we showed previously, flatten can be up to three times slower when the predicate is nonstring_iterable rather than list_or_tuple. We can also write a nonrecursive version of generator flatten. Such a version lets you flatten nested sequences with nesting levels higher than Python's recursion limit, which normally allows no more than a few thousand levels of recursion depth. The main technique for recursion removal is to keep an explicit last-in, first-out (LIFO) stack, which, in this case, we can implement with a list of iterators: def flatten(sequence, to_expand=list_or_tuple): iterators = [ iter(sequence) ] while iterators: # loop on the currently most-nested (last) iterator for item in iterators[-1]: if to_expand(item): # subsequence found, go loop on iterator on subsequence iterators.append(iter(item)) break else: yield item else: # most-nested iterator exhausted, go back, loop on its parent iterators.pop( ) The if clause of the if statement executes for any item we are to expandthat is, any subsequence on which we must loop; so in that clause, we push an iterator for the subsequence to the end of the stack, then execute a break to terminate the for, and go back to the outer while, which will in turn execute a new for statement on the iterator we just appended to the stack. The else clause of the if statement executes for any item we don't expand, and it just yields the item. The else clause of the for statement executes if no break statement interrupts the for loopin other words, when the for loop runs to completion, exhausting the currently most-nested iterator. So, in that else clause, we remove the now-exhausted most-nested (last) iterator, and the outer while loop proceeds, either terminating if no iterators are left on the stack, or executing a new for statement that continues the loop on the iterator that's back at the top of the stackfrom wherever that iterator had last left off, intrinsically, because an iterator's job is exactly to remember iteration state. The results of this nonrecursive implementation of flatten are identical to those of the simpler recursive version given in this recipe's Solution. If you think non-recursive implementations arefaster than recursive ones, though, you may be disappointed: according to my measurements, the nonrecursive version is about 10% slower than the recursive one, across a range of cases. See Also Library Reference and Python in a Nutshell sections on sequence types and built-ins iter, isinstance, and basestring. Recipe 4.7. Removing or Reordering Columnsin a List of Rows Credit: Jason Whitlark Problem You have a list of lists (rows) and need to get another list of the same rows but with some columns removed and/or reordered. Solution A list comprehension works well for this task. Say you have: listOfRows = [ [1,2,3,4], [5,6,7,8], [9,10,11,12] ] You want a list with the same rows but with the second of the four columns removed and the third and fourth ones interchanged. A simple list comprehension that performs this job is: newList = [ [row[0], row[3], row[2]] for row in listOfRows ] An alternative way of coding, that is at least as practical and arguably a bit more elegant, is to use an auxiliary sequence (meaning a list or tuple) that has the column indices you desire in their proper order. Then, you can nest an inner list comprehension that loops on the auxiliary sequence inside the outer list comprehension that loops on listOfRows: newList = [ [row[ci] for ci in (0, 3, 2)] for row in listofRows ] Discussion I often use lists of lists to represent two-dimensional arrays. I think of such lists as having the "rows" of a "two-dimensional array" as their items. I often perform manipulation on the "columns" of such a "two-dimensional array", typically reordering some columns, sometimes omitting some of the original columns. It is not obvious (at least, it was not immediately obvious to me) that list comprehensions are just as useful for this purpose as they are for other kinds of sequence-manipulation tasks. A list comprehension builds a new list, rather than altering an existing one. But even when you do need to alter the existing list in place, the best approach is to write a list comprehension and assign it to the existing list's contents. For example, if you needed to alter listOfRows in place, for the example given in this recipe's Solution, you would code: listOfRows[:] = [ [row[0], row[3], row[2]] for row in listOfRows ] Do consider, as suggested in the second example in this recipe's Solution, the possibility of using an auxiliary sequence to hold the column indices you desire, in the order in which you desire them, rather than explicitly hard-coding the list display as we did in the first example. You mightfeel a little queasy about nesting two list comprehensions into each other in this fashion, but it's simpler and safer than you might fear. If you adopt this approach, you gain some potential generality, because you can choose to give a name to the auxiliary sequence of indices, use it to reorder several lists of rows in the same fashion, pass it as an argument to a function, whatever: def pick_and_reorder_columns(listofRows, column_indexes): return [ [row[ci] for ci in column_indexes] for row in listofRows ] columns = 0, 3, 2 newListOfPandas = pick_and_reorder_columns(oldListOfPandas, columns) newListOfCats = pick_and_reorder_columns(oldListOfCats, columns) This example performs just the same column reordering and selection as all the other snippets in this recipe, but it performs the operation on two separate "old" lists, obtaining from each the corresponding "new" list. Reaching for excessive generalization is a pernicious temptation, but here, with this pick_and_reorder_columns function, it seems that we are probably getting just the right amount of generality. One last note: some people prefer a fancier way to express the kinds of list comprehensions that are used as "inner" ones in some of the functions used previously. Instead of coding them straightforwardly, as in: [row[ci] for ci in column_indexes] they prefer to use the built-in function map, and the special method _ _getitem_ _ of row used as a bound-method, to perform the indexing subtask, so they code instead: map(row._ _getitem_ _, column_indexes) Depending on the exact version of Python, perhaps this fancy and somewhat obscure way may be slightly faster. Nevertheless, I think the greater simplicity of the list comprehension form means the list comprehension is still the best way. See Also List comprehension docs in Language Reference and Python in a Nutshell. Recipe 4.8. Transposing Two-Dimensional Arrays Credit: Steve Holden, Raymond Hettinger, Attila Vàsàrhelyi, Chris Perkins Problem You need to transpose a list of lists, turning rows into columns and vice versa. Solution You must start with a list whose items are lists all of the same length, such as: arr = [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]] A list comprehension offers a simple, handy way to transpose such a two-dimensional array: print [[r[col] for r in arr] for col in range(len(arr[0]))] [[1, 4, 7, 10], [2, 5, 8, 11], [3, 6, 9, 12]] A faster though more obscure alternative (with exactly the same output) can be obtained by exploiting built-in function zip in a slightly strange way: print map(list, zip(*arr)) Discussion This recipe shows a concise yet clear way to turn rows into columns, and also a faster though more obscure way. List comprehensions work well when you want to be clear yet concise, while the alternative solution exploits the built-in function zip in a way that is definitely not obvious. Sometimes data just comes at you the wrong way. For instance, if you use Microsoft's ActiveX Data Ojbects (ADO) database interface, due to array element-ordering differences between Python and Microsoft's preferred implementation language (Visual Basic), the Getrows method actually appears to return database columns in Python, despite the method's name. This recipe's two solutions to this common kind of problem let you choose between clarity and speed. In the list comprehension solution, the inner comprehension varies what is selected from (the row), while the outer comprehension varies the selector (the column). This process achieves the required transposition. In the zip-based solution, we use the *a syntax to pass each item (row) of arr to zip, in order, as a separate positional argument. zip returns a list of tuples, which directly achieves the required transposition; we then apply list to each tuple, via the single call to map, to obtain a list of lists, as required. Since we don't use zip's result as a list directly, we could get a further slight improvement in performance by using itertools.izip instead (because izip does not materialize its result as a list in memory, but rather yields it one item at a time): import itertools print map(list, itertools.izip(*arr))but, in this specific case, the slight speed increase is probably not worth the added complexity. The *args and **kwds Syntax *args (actually, * followed by any identifiermost usually, you'll see args or a as the identifier that's used) is Python syntax for accepting or passing arbitrary positional arguments. When you receive arguments with this syntax (i.e., when you place the star syntax within a function's signature, in the def statement for that function), Python binds the identifier to a tuple that holds all positional arguments not "explicitly" received. When you pass arguments with this syntax, the identifier can be bound to any iterable (in fact, it could be any expression, not necessarily an identifier, as long as the expression's result is an iterable). **kwds (again, the identifier is arbitrary, most often kwds or k) is Python syntax for accepting or passing arbitrary named arguments. (Python sometimes calls named arguments keyword arguments, which they most definitely are notjust try to use as argument name a keyword, such as pass, for, or yield, and you'll see. Unfortunately, this confusing terminology is, by now, ingrained in the language and its culture.) When you receive arguments with this syntax (i.e., when you place the starstar syntax within a function's signature, in the def statement for that function), Python binds the identifier to a dict, which holds all named arguments not "explicitly" received. When you pass arguments with this syntax, the identifier must be bound to a dict (in fact, it could be any expression, not necessarily an identifier, as long as the expression's result is a dict). Whether in defining a function or in calling it, make sure that both *a and **k come after any other parameters or arguments. If both forms appear, then place the **k after the *a. If you're transposing large arrays of numbers, consider Numeric Python and other third-party packages. Numeric Python defines transposition and other axis-swinging routines that will make your head spin. See Also The Reference Manual and Python in a Nutshell sections on list displays (the other name for list comprehensions) and on the *a and *k notation for positional and named argument passing; built-in functions zip and map; Numeric Python (http://www.pfdubois.com/numpy/). Recipe 4.9. Getting a Value from a Dictionary Credit: Andy McKay Problem You need to obtain a value from a dictionary, without having to handle an exception if the key you seek is not in the dictionary. Solution That's what the get method of dictionaries is for. Say you have a dictionary such as d = {'key':'value',}. To get the value corresponding to key in d in an exception-safe way, code: print d.get('key', 'not found') If you need to remove the entry after you have obtained the value, call d.pop (which does a get- and-remove) instead of d.get (which just reads d and never changes it). Discussion Want to get a value for a key from a dictionary, without getting an exception if the key does not exist in the dictionary? Use the simple and useful get method of the dictionary. If you try to get a value with the indexing syntax d[x], and the value of x is not a key in dictionary d, your attempt raises a KeyError exception. This is often okay. If you expected the value of x to be a key in d, an exception is just the right way to inform you that you're mistaken (i.e., that you need to debug your program). However, you often need to be more tentative about it: as far as you know, the value of x may or may not be a key in d. In this case, don't start messing with in tests, such as: if 'key' in d: print d['key'] else: print 'not found' or try/except statements, such as: try: print d['key'] except KeyError: print 'not found' Instead, use the get method, as shown in the "Solution". If you call d.get(x), no exception is thrown: you get d[x] if x is a key in d, and if it's not, you get None (which you can check for or propagate). If None is not what you want to get when x is not a key of d, call d.get(x,somethingelse) instead. In this case, if x is not a key, you will get the value of somethingelse. get is a simple, useful mechanism that is well explained in the Python documentation, but a surprising number of people don't know about it. Another similar method is pop, which is mostly like get, except that, if the key was in the dictionary, pop also removes it. Just one caveat: get and pop are not exactly parallel. d.pop(x) does raise KeyError if x is not a key in d; to get exactly the same effect as d.get(x), plus the entry removal, call d.pop(x,None) instead. See Also Recipe 4.10; the Library Reference and Python in a Nutshell sections on mapping types. Recipe 4.10. Adding an Entry to a Dictionary Credit: Alex Martelli, Martin Miller, Matthew Shomphe Problem Working with a dictionary d, you need to use the entry d[k] when it's already present, or add a new value as d[k] when k isn't yet a key in d. Solution This is what the setdefault method of dictionaries is for. Say we're building a word- to-page- numbers index, a dictionary that maps each word to the list of page numbers where it appears. A key piece of code in that application might be: def addword(theIndex, word, pagenumber): theIndex.setdefault(word, [ ]).append(pagenumber) This code is equivalent to more verbose approaches such as: def addword(theIndex, word, pagenumber): if word in theIndex: theIndex[word].append(pagenumber) else: theIndex[word] = [pagenumber] and: def addword(theIndex, word, pagenumber): try: theIndex[word].append(pagenumber) except KeyError: theIndex[word] = [pagenumber] Using method setdefault simplifies this task considerably. Discussion For any dictionary d, d.setdefault(k, v) is very similar to d.get(k, v), which was covered previously in Recipe 4.9. The essential difference is that, if k is not a key in the dictionary, the setdefault method assigns d[k]=v as a side effect, in addition to returning v. (get would just return v, without affecting d in any way.) Therefore, consider using setdefault any time you have get-like needs, but also want to produce this side effect on the dictionary. setdefault is particularly useful in a dictionary with values that are lists, as detailed in Recipe 4.15. The most typical usage for setdefault is something like: somedict.setdefault(somekey, [ ]).append(somevalue)setdefault is not all that useful for immutable values, such as numbers. If you just want to count words, for example, the right way to code is to use, not setdefault, but rather get: theIndex[word] = theIndex.get(word, 0) + 1 since you must rebind the dictionary entry at theIndex[word] anyway (because numbers are immutable). But for our word-to page-numbers example, you definitely do not want to fall into the performance trap that's hidden in the following approach: def addword(theIndex, word, pagenumber): theIndex[word] = theIndex.get(word, [ ]) + [pagenumber] This latest version of addword builds three new lists each time you call it: an empty list that's passed as the second argument to theIndex.get, a one-item list containing just pagenumber, and a list with N+1 items obtained by concatenating these two (where N is the number of times that word was previously found). Building such a huge number of lists is sure to take its toll, in performance terms. For example, on my machine, I timed the task of indexing the same four words occurring once each on each of 1,000 pages. Taking the first version of addword in the recipe as a reference point, the second one (using try/except) is about 10% faster, the third one (using setdefault) is about 20% slowerthe kind of performance differences that you should blissfully ignore in just about all cases. This fourth version (using get) is four times slowerthe kind of performance difference you just can't afford to ignore. See Also Recipe 4.9; Recipe 4.15; Library Reference and Python in a Nutshell documentation about dict. Recipe 4.11. Building a Dictionary Without Excessive Quoting Credit: Brent Burley, Peter Cogolo Problem You want to construct a dictionary whose keys are literal strings, without having to quote each key. Solution Once you get into the swing of Python, you'll find yourself constructing a lot of dictionaries. When the keys are identifiers, you can avoid quoting them by calling dict with named-argument syntax: data = dict(red=1, green=2, blue=3) This is neater than the equivalent use of dictionary-display syntax: data = {'red': 1, 'green': 2, 'blue': 3} Discussion One powerful way to build a dictionary is to call the built-in type dict. It's often a good alternative to the dictionary-display syntax with braces and colons. This recipe shows that, by calling dict, you can avoid having to quote keys, when the keys are literal strings that happen to be syntactically valid for use as Python identifiers. You cannot use this approach for keys such as the literal strings '12ba' or 'for', because '12ba' starts with a digit, and for happens to be a Python keyword, not an identifier. Also, dictionary-display syntax is the only case in Python where you need to use braces: if you dislike braces, or happen to work on a keyboard that makes braces hard to reach (as all Italian layout keyboards do!), you may be happier, for example, using dict() rather than { } to build an empty dictionary. Calling dict also gives you other possibilities. dict(d) returns a new dictionary that is an independent copy of existing dictionary d, just like d.copy( )but dict(d) works even when d is a sequence of pairs (key, value) instead of being a dictionary (when a key occurs more than once in the sequence, the last appearance of the key applies). A common dictionary-building idiom is: d = dict(zip(the_keys, the_values)) where the_keys is a sequence of keys and the_values a "parallel" sequence of corresponding values. Built-in function zip builds and returns a list of (key, value) pairs, and built-in type dict accepts that list as its argument and constructs a dictionary accordingly. If the sequences are long, it's faster to use module itertools from the standard Python library:import itertools d = dict(itertools.izip(the_keys, the_values)) Built-in function zip constructs the whole list of pairs in memory, while itertools.izip yields only one pair at a time. On my machine, with sequences of 10,000 numbers, the latter idiom is about twice as fast as the one using zip18 versus 45 milliseconds with Python 2.3, 17 versus 32 with Python 2.4. You can use both a positional argument and named arguments in the same call to dict (if the named argument clashes with a key specified in the positional argument, the named argument applies). For example, here is a workaround for the previously mentioned issue that Python keywords, and other nonidentifiers, cannot be used as argument names: d = dict({'12ba':49, 'for': 23}, rof=41, fro=97, orf=42) If you need to build a dictionary where the same value corresponds to each key, call dict.fromkeys(keys_sequence, value) (if you omit the value, it defaults to None). For example, here is a neat way to initialize a dictionary to be used for counting occurrences of various lowercase ASCII letters: import string count_by_letter = dict.fromkeys(string.ascii_lowercase, 0) See Also Library Reference and Python in a Nutshell sections on built-ins dict and zip, and on modules itertools and string. Recipe 4.12. Building a Dict from a List of Alternating Keys and Values Credit: Richard Philips, Raymond Hettinger Problem You want to build a dict from a list of alternating keys and values. Solution The built-in type dict offers many ways to build dictionaries, but not this one, so we need to code a function for the purpose. One way is to use the built-in function zip on extended slices: def dictFromList(keysAndValues): return dict(zip(keysAndValues[::2], keysAndValues[1::2])) A more general approach, which works for any sequence or other iterable argument and not just for lists, is to "factor out" the task of getting a sequence of pairs from a flat sequence into a separate generator. This approach is not quite as concise as dictFromList, but it's faster as well as more general: def pairwise(iterable): itnext = iter(iterable).next while True: yield itnext( ), itnext( ) def dictFromSequence(seq): return dict(pairwise(seq)) Defining pairwise also allows updating an existing dictionary with any sequence of alternating keys and valuesjust code, for example, mydict.update(pairwise(seq)). Discussion Both of the "factory functions" in this recipe use the same underlying way to construct a dictionary: each calls dict with an argument that is a sequence of (key, value) pairs. All the difference is in how the functions build the sequence of pairs to pass to dict. dictFromList builds a list of such pairs by calling built-in function zip with two extended-form slices of the function's keysAndValues argumentone that gathers all items with even indices (meaning the items at index 0, 2, 4, . . .), the other that gathers all items with odd indices (starting at 1 and counting by 2 . . .). This approach is fine, but it works only when the argument named keysAndValues is an instance of a type or class that supports extended slicing, such as list, tuple or str. Also, this approach results in constructing several temporary lists in memory: if keysAndValues is a long sequence, all of this list construction activity can cost some performance. dictFromSequence, on the other hand, delegates the task of building the sequence of pairs to thegenerator named pairwise. In turn, pairwise is coded to ensure that it can use any iterable at allnot just lists (or other sequences, such as tuples or strings), but also, for example, results of other generators, files, dictionaries, and so on. Moreover, pairwise yields pairs one at a time. It never constructs any long list in memory, an aspect that may improve performance if the input sequence is very long. The implementation of pairwise is interesting. As its very first statement, pairwise binds local name itnext to the bound-method next of the iterator that it obtains by calling the built-in function iter on the iterable argument. This may seem a bit strange, but it's a good general technique in Python: if you start with an object, and all you need to do with that object is call one of its methods in a loop, you can extract the bound-method, assign it to a local name, and afterwards just call the local name as if it were a function. pairwise would work just as well if the next method was instead called in a way that may look more normal to programmers who are used to other languages: def pairwise_slow(iterable): it = iter(iterable) while True: yield it.next( ), it.next( ) However, this pairwise_slow variant isn't really any simpler than the pairwise generator shown in the Solution ("more familiar to people who don't know Python" is not a synonym of "simpler"!), and it is about 60% slower. Focusing on simplicity and clarity is one thing, and a very good oneindeed, a core principle of Python. Throwing performance to the winds, without getting any real advantage to compensate, is a completely different proposition and definitely not a practice that can be recommended in any language. So, while it is an excellent idea to focus on writing correct, clear, and simple code, it's also very advisable to learn and use Python's idioms that are most appropriate to your needs. See Also Recipe 19.7 for more general approaches to looping by sliding windows over an iterable. See the Python Reference Manual for more on extended slicing. Recipe 4.13. Extracting a Subset of a Dictionary Credit: David Benjamin Problem You want to extract from a larger dictionary only that subset of it that corresponds to a certain set of keys. Solution If you want to leave the original dictionary intact: def sub_dict(somedict, somekeys, default=None): return dict([ (k, somedict.get(k, default)) for k in somekeys ]) If you want to remove from the original the items you're extracting: def sub_dict_remove(somedict, somekeys, default=None): return dict([ (k, somedict.pop(k, default)) for k in somekeys ]) Two examples of these functions' use and effects: >>> d = {'a': 5, 'b': 6, 'c': 7} >>> print sub_dict(d, 'ab'), d {'a': 5, 'b': 6} {'a': 5, 'b': 6, 'c': 7} >>> print sub_dict_remove(d, 'ab'), d {'a': 5, 'b': 6} {'c': 7} Discussion In Python, I use dictionaries for many purposesdatabase rows, primary and compound keys, variable namespaces for template parsing, and so on. So, I often need to create a dictionary that is based on another, larger dictionary, but only contains the subset of the larger dictionary corresponding to some set of keys. In most use cases, the larger dictionary must remain intact after the extraction; sometimes, however, I need to remove from the larger dictionary the subset that I'm extracting. This recipe's solution shows both possibilities. The only difference is that you use method get when you want to avoid affecting the dictionary that you are getting data from, method pop when you want to remove the items you're getting. If some item k of somekeys is not in fact a key in somedict, this recipe's functions put k as a key in the result anyway, with a default value (which I pass as an optional argument to either function, with a default value of None). So, the result is not necessarily a subset of somedict. This behavior is the one I've found most useful in my applications. You might prefer to get an exception for "missing keys"that would help alert you to a bug in your program, in cases in which you know all ks in somekeys should definitely also be keys in somedict. Remember, "errors should never pass silently. Unless explicitly silenced," to quote TheZen of Python, by Tim Peters (enter the statement import this at an interactive Python prompt to read or re-read this delightful summary of Python's design principles). So, if a missing key is an error, from the point of view of your application, then you do want to get an exception that alerts you to that error at once, if it ever occurs. If this is what you want, you can get it with minor modifications to this recipe's functions: def sub_dict_strict(somedict, somekeys): return dict([ (k, somedict[k]) for k in somekeys ]) def sub_dict_remove_strict(somedict, somekeys): return dict([ (k, somedict.pop(k)) for k in somekeys ]) As you can see, these strict variants are even simpler than the originalsa good indication that Python likes to raise exceptions when unexpected behavior occurs! Alternatively, you might prefer missing keys to be simply omitted from the result. This, too, requires just minor modifications: def sub_dict_select(somedict, somekeys): return dict([ (k, somedict[k]) for k in somekeys if k in somedict]) def sub_dict_remove_select(somedict, somekeys): return dict([ (k, somedict.pop(k)) for k in somekeys if k in somedict]) The if clause in each list comprehension does all we need to distinguish these _select variants from the _strict ones. In Python 2.4, you can use generator expressions, instead of list comprehensions, as the arguments to dict in each of the functions shown in this recipe. Just change the syntax of the calls to dict, from dict([. . .]) to dict(. . .) (removing the brackets adjacent to the parentheses) and enjoy the resulting slight simplification and acceleration. However, these variants would not work in Python 2.3, which has list comprehensions but not generator expressions. See Also Library Reference and Python in a Nutshell documentation on dict. Recipe 4.14. Inverting a Dictionary Credit: Joel Lawhead, Ian Bollinger, Raymond Hettinger Problem An existing dict maps keys to unique values, and you want to build the inverse dict, mapping each value to its key. Solution You can write a function that passes a list comprehension as dict's argument to build the new requested dictionary: def invert_dict(d): return dict([ (v, k) for k, v in d.iteritems( ) ]) For large dictionaries, though, it's faster to use the generator izip from the itertools module in the Python Standard Library: from itertools import izip def invert_dict_fast(d): return dict(izip(d.itervalues( ), d.iterkeys( ))) Discussion If the values in dict d are not unique, then d cannot truly be inverted, meaning that there exists no dict id such that for any valid key k, id[d[k]]==k. However, the functions shown in this recipe still construct, even in such cases, a "pseudo-inverse" dict pd such that, for any v that is a value in d, d[pd[v]]==v. Given the original dict d and the dict x returned by either of the functions shown in this recipe, you can easily check whether x is the true inverse of d or just d's pseudo-inverse: x is the true inverse of d if and only if len(x)==len(d). That's because, if two different keys have the same value, then, in the result of either of the functions in this recipe, one of the two keys will simply go "poof" into the ether, thus leaving the resulting pseudo- inverse dict shorter than the dict you started with. In any case, quite obviously, the functions shown in this recipe can work only if all values in d are hashable (meaning that they are all usable as keys into a dict): otherwise, the functions raise a TypeError exception. When we program in Python, we normally "disregard minor optimizations," as Donald Knuth suggested over thirty years ago: we place a premium on clarity and correctness and care relatively little about speed. However, it can't hurt to know about faster possibilities: when we decide to code in a certain way because it's simpler or clearer than another, it's best if we are taking the decision deliberately, not out of ignorance. Here, function invert_dict in this recipe's Solution might perhaps be considered clearer because it shows exactly what it's doing. Take the pairs k, v of key and value that method iteritems yields, swap them into (value, key) order, and feed the resulting list as the argument of dict, so that dict builds a dictionary where each value v is a key and the corresponding key k becomes that key's valuejust the inverse dict that our problem requires.However, function invert_dict_fast, also in this recipe's Solution, isn't really any more complicated: it just operates more abstractly, by getting all keys and all values as two separate iterators and zipping them up (into an iterator whose items are the needed, swapped (value, key) pairs) via a call to generator izip, supplied by the itertools module of the Python Standard Library. If you get used to such higher abstraction levels, they will soon come to feel simpler than lower-level code! Thanks to the higher level of abstraction, and to never materializing the whole list of pairs (but rather operating via generators and iterators that yield only one item at a time), function invert_dict_fast can be substantially faster than function invert_dict. For example, on my machine, to invert a 10,000-item dictionary, invert_dict takes about 63 milliseconds, but invert_dict_fast manages the same task in just 20 milliseconds. A speed increase by a factor of three, in general, is not to be sneered at. Such performance gains, when you work on large amounts of data, are the norm, rather than the exception, for coding at higher abstraction levels. This is particularly true when you can use itertools rather than loops or list comprehensions, because you don't need to materialize some large list in memory at one time. Performance gain is an extra incentive for getting familiar with working at higher abstraction levels, a familiarity that has conceptual and productivity pluses, too. See Also Documentation on mapping types and itertools in the Library Reference and Python in a Nutshell; Chapter 19. Recipe 4.15. Associating Multiple Values with Each Key in a Dictionary Credit: Michael Chermside Problem You need a dictionary that maps each key to multiple values. Solution By nature, a dictionary is a one-to-one mapping, but it's not hard to make it one-to-manyin other words, to make one key map to multiple values. Your choice of one of two possible approaches depends on how you want to treat duplications in the set of values for a key. The following approach, based on using lists as the dict's values, allows such duplications: d1 = { } d.setdefault(key, [ ]).append(value) while an alternative approach, based on using sub-dicts as the dict's values, automatically eliminates duplications of values: d2 = { } d2.setdefault(key, { })[value] = 1 In Python 2.4, the no-duplication approach can equivalently be coded: d3 = { } d3.setdefault(key, set( )).add(value) Discussion A normal dictionary performs a simple mapping of each key to one value. This recipe shows three easy, efficient ways to achieve a mapping of each key to multiple values, by holding as the dictionary's values lists, sub-dicts, or, in Python 2.4, sets. The semantics of the list-based approach differ slightly but importantly from those of the other two in terms of how they deal with duplication. Each approach relies on the setdefault method of a dictionary, covered earlier in Recipe 4.10, to initialize the entry for a key in the dictionary, if needed, and in any case to return said entry. You need to be able to do more than just add values for a key. With the first approach, which uses lists and allows duplications, here's how to retrieve the list of values for a key: list_of_values = d1[key] Here's how to remove one value for a key, if you don't mind leaving empty lists as items of d1 when the last value for a key is removed:d1[key].remove(value) Despite the empty lists, it's still easy to test for the existence of a key with at least one valuejust use a function that always returns a list (maybe an empty one), such as: def get_values_if_any(d, key): return d.get(key, [ ]) For example, to check whether 'freep' is among the values (if any) for key 'somekey' in dictionary d1, you can code: if 'freep' in get_values_if_any(d1, 'somekey'). The second approach, which uses sub-dicts and eliminates duplications, can use rather similar idioms. To retrieve the list of values for a key: list_of_values = list(d2[key]) To remove one value for a key, leaving empty dictionaries as items of d2 when the last value for a key is removed: del d2[key][value] In the third approach, showing the Python 2.4-only version d3, which uses sets, this would be: d3[key].remove(value) One possibility for the get_values_if_any function in either the second or third (duplication- removing) approaches would be: def get_values_if_any(d, key): return list(d.get(key, ( ))) This recipe focuses on how to code the raw functionality, but, to use this functionality in a systematic way, you'll probably want to wrap up this code into a class. For that purpose, you need to make some of the design decisions that this recipe highlights. Do you want a value to be in the entry for a key multiple times? (Is the entry for each key a bag rather than a set, in mathematical terms?) If so, should remove just reduce the number of occurrences by 1, or should it wipe out all of them? This is just the beginning of the choices you have to make, and the right choices depend on the specifics of your application. See Also Recipe 4.10; the Library Reference and Python in a Nutshell sections on mapping types; Recipe 18.8 for an implementation of the bag type. Recipe 4.16. Using a Dictionary to Dispatch Methods or Functions Credit: Dick Wall Problem You need to execute different pieces of code depending on the value of some control variablethe kind of problem that in some other languages you might approach with a case statement. Solution Object-oriented programming, thanks to its elegant concept of dispatching, does away with many (but not all) needs for case statements. In Python, dictionaries, and the fact that functions are first-class objects (in particular, functions can be values in a dictionary), conspire to make the full problem of "case statements" easier to solve. For example, consider the following snippet of code: animals = [ ] number_of_felines = 0 def deal_with_a_cat( ): global number_of_felines print "meow" animals.append('feline') number_of_felines += 1 def deal_with_a_dog( ): print "bark" animals.append('canine') def deal_with_a_bear( ): print "watch out for the *HUG*!" animals.append('ursine') tokenDict = { "cat": deal_with_a_cat, "dog": deal_with_a_dog, "bear": deal_with_a_bear, } # Simulate, say, some words read from a file words = ["cat", "bear", "cat", "dog"] for word in words: # Look up the function to call for each word, and call it return tokenDict[word]( ) nf = number_of_felines print 'we met %d feline%s' % (nf, 's'[nf==1:]) print 'the animals we met were:', ' '.join(animals) Discussion The key idea in this recipe is to construct a dictionary with string (or other) values as keys, and bound-methods, functions, or other callables as values. At each step of execution, we use thestring keys to select which callable to execute and then call it. This approach can be used as a kind of generalized case statement. It's embarrassingly simple (really!), but I use this technique often. You can also use bound- methods or other callables instead of functions. If you use unbound methods, you need to pass an appropriate object as the first actual argument when you do call them. More generally, you can store, as the dictionary's values, tuples including both a callable and arguments to pass to the callable. I primarily use this technique in places where in other languages, I might want a case, switch, or select statement. For example, I use it to implement a poor man's way to parse command files (e.g., an X10 macro control file). See Also The Library Reference section on mapping types; the Reference Manual section on bound and unbound methods; Python in a Nutshell about both dictionaries and callables. Recipe 4.17. Finding Unions and Intersections of Dictionaries Credit: Tom Good, Andy McKay, Sami Hangaslammi, Robin Siebler Problem Given two dictionaries, you need to find the set of keys that are in both dictionaries (the intersection) or the set of keys that are in either dictionary (the union). Solution Sometimes, particularly in Python 2.3, you find yourself using dictionaries as concrete representations of sets. In such cases, you only care about the keys, not the corresponding values, and often you build the dictionaries by calls to dict.fromkeys, such as a = dict.fromkeys(xrange(1000)) b = dict.fromkeys(xrange(500, 1500)) The fastest way to compute the dict that is the set-union is: union = dict(a, **b) The fastest concise way to compute the dict that is the set-intersection is: inter = dict.fromkeys([x for x in a if x in b]) If the number of items in dictionaries a and b can be very different, then it can be important for speed considerations to have the shorter one in the for clause, and the longer one in the if clause, of this list comprehension. In such cases, it may be worth sacrificing some conciseness in favor of speed, by coding the intersection computation as follows: if len(a) < len(b): inter = dict.fromkeys([x for x in a if x not in b]) else: inter = dict.fromkeys([x for x in b if x not in a]) Python also gives you types to represent sets directly (in standard library module sets, and, in Python 2.4, also as built-ins). Here is a snippet that you can use at the start of a module: the snippet ensures that name set is bound to the best available set type, so that throughout the module, you can then use the same code whether you're using Python 2.3 or 2.4: try: set except NameError: from sets import Set as setHaving done this, you can now use type set to best effect, gaining clarity and conciseness, and (in Python 2.4) gaining a little speed, too: a = set(xrange(1000)) b = set(xrange(500, 1500)) union = a | b inter = a & b Discussion In Python 2.3, even though the Python Standard Library module sets offers an elegant data type Set that directly represents a set (with hashable elements), it is still common to use a dict to represent a set, partly for historical reasons. Just in case you want to keep doing it, this recipe shows you how to compute unions and intersections of such sets in the fastest ways, which are not obvious. The code in this recipe, on my machine, takes about 260 microseconds for the union, about 690 for the intersection (with Python 2.3; with Python 2.4, 260 and 600,respectively), while alternatives based on loops or generator expressions are substantially slower. However, it's best to use type set instead of representing sets by dictionaries. As the recipe shows, using set makes your code more direct and readable. If you dislike the or-operator (|) and the "and-operator" (&), you can equivalently use a.union(b) and a.intersection(b), respectively. Besides clarity, you also gain speed, particularly in Python 2.4: computing the union still takes about 260 microseconds, but computing the intersection takes only about 210. Even in Python 2.3, this approach is acceptably fast: computing the union takes about 270 microseconds, computing the intersection takes about 650not quite as fast as Python 2.4 but still quite comparable to what you can get if you represent sets by dictionaries. Last but not least, once you use type set (whether it is the Python 2.4 built-in, or class Set from the Python Standard Library module sets, the interface is the same), you gain a wealth of useful set operations. For example, the set of elements that are in either a or b but not both is a^b or, equivalently, a.symmetric_difference(b). Even if you start with dicts for other reasons, consider using sets anyway if you need to perform set operations. Say, for example, that you have in phones a dictionary that maps names to phone numbers and in addresses one that maps names to addresses. The clearest and simplest way to print all names for which you know both address and phone number, and their associated data, is: for name in set(phones) & set(addresses): print name, phones[name], addresses[name] This is much terser, and arguably clearer, than something like: for name in phones: if name in addresses: print name, phones[name], addresses[name] Another excellent alternative is: for name in set(phones).intersection(addresses): print name, phones[name], addresses[name] If you use the named intersection method, rather than the & intersection operator, you don't need to turn both dicts into sets: just one of them. Then call intersection on the resulting set, and pass the other dict as the argument to the intersection method.See Also The Library Reference and Python in a Nutshell sections on mapping types, module sets, and Python 2.4's built-in set type. Recipe 4.18. Collecting a Bunch of Named Items Credit: Alex Martelli, Doug Hudgeon Problem You want to collect a bunch of items together, naming each item of the bunch, and you find dictionary syntax a bit heavyweight for the purpose. Solution Any normal class instance inherently wraps a dictionary, which it uses to hold its state. We can easily take advantage of this handily wrapped dictionary by coding a nearly empty class: class Bunch(object): def _ _init_ _(self, **kwds): self._ _dict_ _.update(kwds) Now, to group a few variables, create a Bunch instance: point = Bunch(datum=y, squared=y*y, coord=x) You can now access and rebind the named attributes just created, add others, remove some, and so on. For example: if point.squared > threshold: point.isok = True Discussion We often just want to collect a bunch of stuff together, naming each item of the bunch. A dictionary is OK for this purpose, but a small do-nothing class is even handier and prettier to use. It takes minimal effort to build a little class, as in this recipe, to provide elegant attribute-access syntax. While a dictionary is fine for collecting a few items in which each item has a name (the item's key in the dictionary can be thought of as the item's name, in this context), it's not the best solution when all names are identifiers, to be used just like variables. In class Bunch's _ _init_ _ method, we accept arbitrary named arguments with the **kwds syntax, and we use the kwds dictionary to update the initially empty instance dictionary, so that each named argument gets turned into an attribute of the instance. Compared to attribute-access syntax, dictionary-indexing syntax is not quite as terse and readable. For example, if point was a dictionary, the little snippet at the end of the "Solution" would have to be coded like: if point['squared'] > threshold: point['isok'] = TrueAn alternative implementation that's just as attractive as the one used in this recipe is: class EvenSimplerBunch(object): def _ _init_ _(self, **kwds): self._ _dict_ _ = kwds Rebinding an instance's dictionary may feel risqué, but it's not actually any pushier than calling that dictionary's update method. So you might prefer the marginal speed advantage of this alternative implementation of Bunch. Unfortunately, I cannot find anywhere in Python's documentation an assurance that usage like: d = {'foo': 'bar'} x = EvenSimplerBunch(**d) will forever keep making x._ _dict_ _ an independent copy of d rather than just sharing a reference. It does currently, and in every version, but unless it's a documented semantic constraint, we cannot be entirely sure that it will keep working forever. So, if you do choose the implementation in EvenSimplerBunch, you might choose to assign a copy (dict(kwds) or kwds.copy( )) rather than kwds itself. And, if you do, then the marginal speed advantage disappears. All in all, the Bunch presented in this recipe's Solution is probably preferable. A further tempting but not fully sound alternative is to have the Bunch class inherit from dict, and set attribute access special methods equal to the item access special methods, as follows: class DictBunch(dict): _ _getattr_ _ = dict._ _getitem_ _ _ _setattr_ _ = dict._ _setitem_ _ _ _delattr_ _ = dict._ _delitem_ _ One problem with this approach is that, with this definition, an instance x of DictBunch has many attributes it doesn't really have, because it inherits all the attributes (methods, actually, but there's no significant difference in this context) of dict. So, you can't meaningfully check hasattr(x, someattr), as you could with the classes Bunch and EvenSimplerBunch previously shown, unless you can somehow rule out the value of someattr being any of several common words such as 'keys', 'pop', and 'get'. Python's distinction between attributes and items is really a wellspring of clarity and simplicity. Unfortunately, many newcomers to Python wrongly believe that it would be better to confuse items with attributes, generally because of previous experience with JavaScript and other such languages, in which attributes and items are regularly confused. But educating newcomers is a much better idea than promoting item/attribute confusion. See Also The Python Tutorial section on classes; the Language Reference and Python in a Nutshell coverage of classes; Chapter 6 for more information about object-oriented programming in Python; Recipe 4.18 for more on the **kwds syntax. Recipe 4.19. Assigning and Testing with One Statement Credit: Alex Martelli, Martin Miller Problem You are transliterating C or Perl code to Python, and to keep close to the original's structure, you'd like an expression's result to be both assigned and tested (as in if((x=foo( )) or while((x=foo( )) in such other languages). Solution In Python, you can't code if x=foo(): . . . . Assignment is a statement, so it cannot fit into an expression, and you can only use expressions as conditions of if and while statements. This isn't a problem, it just means you have to structure your code Pythonically! For example, to process a file object f line by line, instead of the following C-like (and syntactically incorrect, in Python) approach: while (line=f.readline( )) != '': process(line) you can code a highly Pythonic (readable, clean, fast) approach: for line in f: process(line) But sometimes, you're transliterating from C, Perl, or another language, and you'd like your transliteration to be structurally close to the original. One simple utility class makes it easy: class DataHolder(object): def _ _init_ _(self, value=None): self.value = value def set(self, value): self.value = value return value def get(self): return self.value # optional and strongly discouraged, but nevertheless handy at times: import _ _builtin_ _ _ _builtin_ _.DataHolder = DataHolder _ _builtin_ _.data = data = DataHolder( ) With the help of the DataHolder class and its instance data, you can keep your C-like code structure intact in transliteration: while data.set(file.readline( )) != '': process(data.get( ))Discussion In Python, assignment is a statement, not an expression. Thus, you cannot assign the result that you are also testing, for example, in the condition of an if, elif, or while statement. This is usually fine: just structure your code to avoid the need to assign while testing (in fact, your code will often become clearer as a result). In particular, whenever you feel the need to assign-and- test within the condition of a while loop, that's a good hint that your loop's structure probably wants to be refactored into a generator (or other iterator). Once you have refactored in this way, your loops become plain and simple for statements. The example given in the recipe, looping over each line read from a text file, is one where the refactoring has already been done on your behalf by Python itself, since a file object is an iterator whose items are the file's lines. However, sometimes you may be writing Python code that is the transliteration of code originally written in C, Perl, or some other language that supports assignment-as-expression. Such transliterations often occur in the first Python version of an algorithm for which a reference implementation is supplied, an algorithm taken from a book, and so on. In such cases, it's often preferable to have the structure of your initial transliteration be close to that of the code you're transcribing. You can refactor later and make your code more Pythonicclearer, faster, and so on. But first, you want to get working code as soon as possible, and specifically you want code that is easy to check for compliance to the original it has been transliterated from. Fortunately, Python offers enough power to make it quite easy for you to satisfy this requirement. Python doesn't let us redefine the meaning of assignment, but we can have a method (or function) that saves its argument somewhere and also returns that argument so it can be tested. That somewhere is most naturally an attribute of an object, so a method is a more natural choice than a function. Of course, we could just retrieve the attribute directly (i.e., the get method is redundant), but it looks nicer to me to have symmetry between data.set and data.get. data.set(whatever) can be seen as little more than syntactic sugar around data.value=whatever, with the added value of being acceptable as an expression. Therefore, it's the one obviously right way to satisfy the requirement for a reasonably faithful transliteration. The only difference between the resulting Python code and the original (say) C or Perl code, is at the syntactic sugar levelthe overall structure is the same, and that's the key issue. Importing _ _builtin_ _ and assigning to its attributes is a trick that basically defines a new built-in object at runtime. You can use that trick in your application's start-up code, and then all other modules will automatically be able to access your new built-ins without having to do an import. It's not good Python practice, though; on the contrary, it's pushing the boundaries of Pythonic good taste, since the readers of all those other modules should not have to know about the strange side effects performed in your application's startup code. But since this recipe is meant to offer a quick-and-dirty approach for a first transliteration that will soon be refactored to make it better, it may be acceptable in this specific context to cut more corners than one would in production-level code. On the other hand, one trick you should definitely not use is the following abuse of a currently existing wart in list comprehensions: while [line for line in [f.readline( )] if line!='']: process(line) This trick currently works, since both Python 2.3 and 2.4 still "leak" the list comprehension control variable (here, line) into the surrounding scope. However, besides being obscure and unreadable, this trick is specifically deprecated: list comprehension control variable leakage will be fixed in some future version of Python, and this trick will then stop working at all. See Also The Tutorial section on classes; the documentation for the _ _builtin_ _ module in the LibraryReference and Python in a Nutshell; Language Reference and Python in a Nutshell documentation on list comprehensions. Recipe 4.20. Using printf in Python Credit: Tobias Klausmann, Andrea Cavalcanti Problem You'd like to output something to your program's standard output with C's function printf, but Python doesn't have that function. Solution It's easy to code a printf function in Python: import sys def printf(format, *args): sys.stdout.write(format % args) Discussion Python separates the concepts of output (the print statement) and formatting (the % operator), but if you prefer to have these concepts together, they're easy to join, as this recipe shows. No more worries about automatic insertion of spaces or newlines, either. Now you need worry only about correctly matching format and arguments! For example, instead of something like: print 'Result tuple is: %r' % (result_tuple,), with its finicky need for commas in unobvious places (i.e., one to make a singleton tuple around result_tuple, one to avoid the newline that print would otherwise insert by default), once you have defined this recipe's printf function, you can just write: printf('Result tuple is: %r', result_tuple) See Also Library Reference and Python in a Nutshell documentation for module sys and for the string formatting operator %; Recipe 2.13 for a way to implement C++'s <<-style output in Python. Recipe 4.21. Randomly Picking Items with Given Probabilities Credit: Kevin Parks, Peter Cogolo Problem You want to pick an item at random from a list, just about as random.choice does, but you need to pick the various items with different probabilities given in another list, rather than picking any item with equal probability as random.choice does. Solution Module random in the standard Python library offers a wealth of possibilities for generating and using pseudo-random numbers, but it does not offer this specific functionality, so we must code it as a function of our own: import random def random_pick(some_list, probabilities): x = random.uniform(0, 1) cumulative_probability = 0.0 for item, item_probability in zip(some_list, probabilities): cumulative_probability += item_probability if x < cumulative_probability: break return item Discussion Module random in the standard Python library does not have the weighted choice functionality that is sometimes needed in games, simulations, and random tests, so I wrote this recipe to supply this functionality. The recipe uses module random's function uniform to get a uniformly distributed pseudo-random number between 0.0 and 1.0, then loops in parallel on items and their probabilities, computing the increasing cumulative probability, until the latter becomes greater than the pseudo-random number. The recipe assumes, but does not check, that probabilities is a sequence with just as many items as some_list, which are probabilitiesthat is, numbers between 0.0 and 1.0, summing up to 1.0; if these assumptions are violated, you may still get some random picks, but they will not follow the (inconsistent) specifications encoded in the function's arguments. You may want to add some assert statements at the start of the function to check that the arguments make sense, such as: assert len(some_list) == len(probabilities) assert 0 <= min(probabilities) and max(probabilities) <= 1 assert abs(sum(probabilities)-1.0) < 1.0e-5 However, these checks can be quite time consuming, so I don't normally use them and have not included them in the official Solution. As I already mentioned, the problem solved in this recipe requires items to be associated withprobabilitiesnumbers between 0 and 1, summing up to 1. A related but slightly different task is to get random picks with weighted relative probabilities given by small non-negative integersodds, rather than probabilities. For this related problem, the best solution is a generator, with an internal structure that is rather different from the function random_pick given in this recipe's Solution: import random def random_picks(sequence, relative_odds): table = [ z for x, y in zip(sequence, relative_odds) for z in [x]*y ] while True: yield random.choice(table) This generator works by first preparing a table whose total number of items is sum(relative_odds), each item of seq appearing in the table as many times as the small non- negative integer that is its corresponding item in relative_odds. Once the table is prepared, the generator's body is tiny and fast, as it simply delegates to random.choice the picking of each random item it yields. Typical uses of this random_picks generator might be: >>> x = random_picks('ciao', [1, 1, 3, 2]) >>> for two_chars in zip('boo', x): print ''.join(two_chars), bc oa oa >>> import itertools >>> print ''.join(itertools.islice(x, 8)) icacaoco See Also Module random in the Library Reference and Python in a Nutshell. Recipe 4.22. Handling Exceptions Within an Expression Credit: Chris Perkins, Gregor Rayman, Scott David Daniels Problem You want to code an expression, so you can't directly use the statement TRy/except, but you still need to handle exceptions that the expression may throw. Solution To catch exceptions, TRy/except is indispensable, and, since try/except is a statement, the only way to use it inside an expression is to code an auxiliary function: def throws(t, f, *a, **k): '''Return True iff f(*a, **k) raises an exception whose type is t (or, one of the items of _tuple_ t, if t is a tuple).''' try: f(*a, **k) except t: return True else: return False For example, suppose you have a text file, which has one number per line, but also extra lines which may be whitespace, comments, or what-have-you. Here is how you can make a list of all the numbers in the file, skipping the lines that aren't numbers: data = [float(line) for line in open(some_file) if not throws(ValueError, float, line)] Discussion You might prefer to name such a function raises, but I personally prefer throws, which is probably a throwback to C++. By whatever name, the auxiliary function shown in this recipe takes as its arguments, first an exception type (or tuple of exception types) t, then a callable f, and then arbitrary positional and named arguments a and k, which are to be passed on to f. Do not code, for example, if not throws(ValueError, float(line))! When you call a function, Python evaluates the arguments before passing control to the function; if an argument's evaluation raises an exception, the function never even gets started. I've seen this erroneous usage attempted more than once by people who are just starting to use the assertRaises method from the standard Python library's unittest.TestCase class, for example. When throws executes, it just calls f within the try clause of a try/except statement, passing on the arbitrary positional and named arguments. If the call to f in the try clause raises an exception whose type is t (or one of the items of t, if t is a tuple of exception types), then control passes to the corresponding except clause, which, in this case, returns true as throws' result. If no exception is raised in the try clause, then control passes to the corresponding else clause (if any), which, in this case, returns False as throws' result.Note that, if some unexpected exception (one whose type is not in t) gets raised, then function throws does not catch that exception, so that throws terminates and propagates the exception to its caller. This choice is quite a deliberate one. Catching exceptions with a too-wide except clause is a bug-diagnosing headache waiting to happen. If the caller really wants throws to catch just about everything, it can always call tHRows(Exception, . . .and live with the resulting headaches. One problem with the throws function is that you end up doing the key operation twiceonce just to see if it throws, tossing the result away, then, a second time, to get the result. It would be nicer to get the result, if any, together with an indication of whether an exception has been caught. I first tried something along the lines of: def throws(t, f, *a, **k): " Return a pair (True, None) if f(*a, **k) raises an exception whose type is in t, else a pair (False, x) where x is the result of f(*a, **k). " try: return False, f(*a, **k) except t: return True, None Unfortunately, this version doesn't fit in well in a list comprehension: there is no elegant way to get and use both the flag and the result. So, I chose a different approach: a function that returns a list in any caseempty if an exception was caught, otherwise with the result as the only item. This approach works fine in a list comprehension, but for clarity, the name of the function needs to be changed: def returns(t, f, *a, **k): " Return [f(*a, **k)] normally, [ ] if that raises an exception in t. " try: return [ f(*a, **k) ] except t: return [ ] The resulting list comprehension is even more elegant, in my opinion, than the original one in this recipe's Solution: data = [ x for line in open(some_file) for x in returns(ValueError, float, line) ] See Also Python in a Nutshell's section on catching and handling exceptions; the sidebar The *args and **kwds Syntax for an explanation of *args and **kwds syntax. Recipe 4.23. Ensuring a Name Is Defined in a Given Module Credit: Steven Cummings Problem You want to ensure that a certain name is defined in a given module (e.g., you want to ensure that there is a built-in name set), and, if not, you want to execute some code that sets the definition. Solution The solution to this problem is the only good use I've yet seen for statement exec. exec lets us execute arbitrary Python code from a string, and thus lets us write a very simple function to deal with this task: import _ _builtin_ _ def ensureDefined(name, defining_code, target=_ _builtin_ _): if not hasattr(target, name): d = { } exec defining_code in d assert name in d, 'Code %r did not set name %r' % ( defining_code, name) setattr(target, name, d[name]) Discussion If your code supports several versions of Python (or of some third-party package), then many of your modules must start with code such as the following snippet (which ensures name set is properly set in either Python 2.4, where it's a built-in, or 2.3, where it must be obtained from the standard library): try: set except NameError: from sets import Set as set This recipe encapsulates this kind of logic directly, and by default works on module _ _builtin_ _, since that's the typical module for which you need to work around missing names in older Python versions. With this recipe, you could ensure name set is properly defined among the built-ins by running just once, during your program's initialization, the single call: ensureDefined('set', 'from sets import Set as set') The key advantage of this recipe is that you can group all needed calls to ensureDefined in just one place of your application, at initialization time, rather than having several ad hoc try/except statements at the start of various modules. Moreover, ensureDefined may allow more readablecode because it does only one specific job, so the purpose of calling it is obvious, while try/except statements could have several purposes, so that more study and reflection might be needed to understand them. Last but not least, using this recipe lets you avoid the warnings that the try/except approach can trigger from such useful checking tools as pychecker, http://pychecker.sourceforge.net/. (If you aren't using pychecker or something like that, you should!) The recipe takes care to avoid unintended accidental side effects on target, by using an auxiliary dictionary d as the target for the exec statement and then transferring only the requested name. This way, for example, you can use as target an object that is not a module (a class, say, or even a class instance), without necessarily adding to your target an attribute named _ _builtins_ _ that references the dictionary of Python's built-ins. If you used less care, so that the body of the if statement was only: exec defining_code in vars(target) you would inevitably get such side effects, as documented at http://www.python.org/doc/current/ref/exec.html. It's important to be aware that exec can and does execute any valid string of Python code that you give it. Therefore, make sure that the argument defining_code that you pass to any call of function ensureDefined does not come from an untrusted source, such as a text file that might have been maliciously tampered with. See Also The online documentation of the exec statement in the Python Language Reference Manual at http://www.python.org/doc/current/ref/exec.html. Chapter 5. Searching and Sorting Introduction Recipe 5.1. Sorting a Dictionary Recipe 5.2. Sorting a List of Strings Case-Insensitively Recipe 5.3. Sorting a List of Objects by an Attribute of the Objects Recipe 5.4. Sorting Keys or Indices Basedon the Corresponding Values Recipe 5.5. Sorting Strings with Embedded Numbers Recipe 5.6. Processing All of a List's Items in Random Order Recipe 5.7. Keeping a Sequence Ordered as Items Are Added Recipe 5.8. Getting the First Few Smallest Items of a Sequence Recipe 5.9. Looking for Items in a Sorted Sequence Recipe 5.10. Selecting the nth Smallest Element of a Sequence Recipe 5.11. Showing off quicksort in Three Lines Recipe 5.12. Performing Frequent Membership Tests on a Sequence Recipe 5.13. Finding Subsequences Recipe 5.14. Enriching the Dictionary Type with Ratings Functionality Recipe 5.15. Sorting Names and Separating Them by Initials Introduction Credit: Tim Peters, PythonLabs Computer manufacturers of the 1960s estimated that more than 25 percent of the running time on their computers was spent on sorting, when all their customers were taken into account. In fact, there were many installations in which the task of sorting was responsible for more than half of the computing time. From these statistics we may conclude that either (i) there are many important applications of sorting, or (ii) many people sort when they shouldn't, or (iii) inefficient sorting algorithms have been in common use. Donald Knuth The Art of Computer Programming,vol. 3, Sorting and Searching, page 3 Professor Knuth's masterful work on the topics of sorting and searching spans nearly 800 pages of sophisticated technical text. In Python practice, we reduce it to two imperatives (we read Knuth so you don't have to): When you need to sort, find a way to use the built-in sort method of Python lists. When you need to search, find a way to use built-in dictionaries. Many recipes in this chapter illustrate these principles. The most common theme is using the decorate-sort-undecorate (DSU) pattern, a general approach to transforming a sorting problem by creating an auxiliary list that we can then sort with the default, speedy sort method. This technique is the single most useful one to take from this chapter. In fact, DSU is so useful that Python 2.4 introduced new features to make it easier to apply. Many recipes can be made simpler in 2.4 as a result, and the discussion of older recipes have been updated to show how. DSU relies on an unusual feature of Python's built-in comparisons: sequences are compared lexicographically. Lexicographical order is a generalization to tuples and lists of the everyday rules used to compare strings (e.g., alphabetical order). The built-in cmp(s1, s2), when s1 and s2 are sequences, is equivalent to this Python code: def lexcmp(s1, s2): # Find leftmost nonequal pair. i = 0 while i < len(s1) and i < len(s2): outcome = cmp(s1[i], s2[i]) if outcome: return outcome i += 1 # All equal, until at least one sequence was exhausted. return cmp(len(s1), len(s2)) This code looks for the first unequal corresponding elements. If such an unequal pair is found, that pair determines the outcome. Otherwise, if one sequence is a proper prefix of the other, the prefix is considered to be the smaller sequence. Finally, if these cases don't apply, the sequences are identical and are considered equal. Here are some examples: >>> cmp((1, 2, 3), (1, 2, 3)) # identical 0 >>> cmp((1, 2, 3), (1, 2)) # first larger because second is a prefix 1>>> cmp((1, 100), (2, 1)) # first smaller because 1<2 -1 >>> cmp((1, 2), (1, 3)) # first smaller because 1==1, then 2<3 -1 An immediate consequence of lexicographical comparison is that if you want to sort a list of objects by a primary key, breaking ties by comparing a secondary key, you can simply build a list of tuples, in which each tuple contains the primary key, secondary key, and original object, in that order. Because tuples are compared lexicographically, this automatically does the right thing. When comparing tuples, the primary keys are compared first, and if (and only if) the primary keys are equal, the secondary keys are compared. The examples of the DSU pattern in this chapter show many applications of this idea. The DSU technique applies to any number of keys. You can add to the tuples as many keys as you like, in the order in which you want the keys compared. In Python 2.4, you can get the same effect with the new key= optional argument to sort, as several recipes point out. Using the sort method's key= argument is easier, more memory-efficient, and runs faster than building an auxiliary list of tuples by hand. The other 2.4-introduced innovation in sorting is a convenient shortcut: a sorted built-in function that sorts any iterable, not in-place, but by first copying it into a new list. In Python 2.3 (apart from the new optional keyword arguments, which apply to the sorted built-in function as well as to list.sort), you can code the same functionality quite easily: def sorted_2_3(iterable): alist = list(iterable) alist.sort( ) return alist Because copying a list and sorting it are both nontrivial operations, and the built-in sorted needs to perform those operations too, no speed advantage is gained in making sorted a built-in. Its advantage is just the convenience. Having something always around and available, rather than having to code even just four simple lines over and over, does make a difference in practice. On the other hand, few tiny functions are used commonly enough to justify expanding the set of built-ins. Python 2.4 added sorted and reversed because those two functions were requested very frequently over the years. The biggest change in Python sorting since the first edition of this book is that Python 2.3 moved to a new implementation of sorting. The primary visible consequences are increased speed in many common cases, and the fact that the new sort is stable (meaning that when two elements compare equal in the original list, they retain their relative order in the sorted list). The new implementation was so successful, and the chances of improving on it appeared so slim, that Guido was persuaded to proclaim that Python's list.sort method will always be stable. This guarantee started with Python 2.4 but was actually realized in Python 2.3. Still, the history of sorting cautions us that better methods may yet be discovered. A brief account of Python's sorting history may be instructive in this regard. A Short History of Python Sorting In early releases of Python, list.sort used the qsort routine from the underlying platform's C library. This didn't work out for several reasons, primarily because the quality of qsort varied widely across machines. Some versions were extremely slow when given a list with many equal values or in reverse-sorted order. Some versions even dumped core because they weren't reentrant. A user-defined _ _cmp_ _ function can also invoke list.sort, so that one list.sort can invoke others as a side effect of comparing. Some platform qsort routines couldn't handle that. A user-defined _ _cmp_ _ function can also (if it's insane or malicious) mutate the list while it's being sorted, and many platform qsort routines dumped core when that happened.Python then grew its own implementation of the quicksort algorithm. This was rewritten with every release, as real-life cases of unacceptable slowness were discovered. Quicksort is a delicate algorithm indeed! In Python 1.5.2 the quicksort algorithm was replaced by a hybrid of samplesort and binary insertion sort, and that implementation remained unchanged for more than four years, until Python 2.3. Samplesort can be viewed as a variant of quicksort that uses a very large sample size to pick the partitioning element, also known as the pivot (it recursively samplesorts a large random subset of the elements and picks the median of those). This variant makes quadratic- time behavior almost impossible and brings the number of comparisons in the average case much closer to the theoretical minimum. However, because samplesort is a complicated algorithm, it has too much administrative overhead for small lists. Therefore, small lists (and small slices resulting from samplesort partitioning) were handled by a separate binary insertion sort, which is an ordinary insertion sort, except that it uses binary search to determine where each new element belongs. Most sorting texts say this isn't worth the bother, but that's because most texts assume that comparing two elements is as cheap as or cheaper than swapping them in memory, which isn't true for Python's sort! Moving an object is very cheap, since what is copied is just a reference to the object. Comparing two objects is expensive, though, because all of the object-oriented machinery for finding the appropriate code to compare two objects and for coercion gets reinvoked each time. This made binary search a major win for Python's sort. On top of this hybrid approach, a few common special cases were exploited for speed. First, already-sorted or reverse-sorted lists were detected and handled in linear time. For some applications, these kinds of lists are very common. Second, if an array was mostly sorted, with just a few out-of-place elements at the end, the binary insertion sort handled the whole job. This was much faster than letting samplesort have at it and occurred often in applications that repeatedly sort a list, append a few new elements, then sort it again. Finally, special code in the samplesort looked for stretches of equal elements, so that the slice they occupy could be marked as done early. In the end, all of this yielded an in-place sort with excellent performance in all known real cases and supernaturally good performance in some common special cases. It spanned about 500 lines of complicated C code, which gives special poignancy to recipe Recipe 5.11. Over the years samplesort was in use, I made a standing offer to buy dinner for anyone who could code a faster Python sort. Alas, I ate alone. Still, I kept my eye on the literature because several aspects of the samplesort hybrid were irritating: While no case of quadratic-time behavior appeared in real life, I knew such cases could be contrived, and it was easy to contrive cases two or three times slower than average ones. The special cases to speed sorting in the presence of extreme partial order were valuable in practice, but my real data often had many other kinds of partial order that should be exploitable. In fact, I came to believe that random ordering in input lists almost never exists in real life (i.e., not outside of timing harnesses for testing sorting algorithms!). There is no practical way to make samplesort stable without grossly increasing memory use. The code was very complex and complicated in ugly ways by the special cases. Current Sorting It was always clear that a mergesort would be better on several counts, including guaranteed worst-case n log n time, and that mergesort is easy to make stable. The problem was that half a dozen attempts to code a mergesort for Python yielded a sort that ran slower (mergesort doesmuch more data movement than samplesort) and consumed more memory. A large and growing literature concerns adaptive sorting algorithms, which attempt to detect order of various kinds in the input. I coded a dozen of them, but they were all much slower than Python's samplesort except on the cases they were designed to exploit. The theoretical bases for these algorithms were simply too complex to yield effective practical algorithms. Then I read an article pointing out that list merging naturally reveals many kinds of partial order, simply by paying attention to how often each input list "wins" in a row. This information was simple and general. When I realized how it could be applied to a natural mergesort, which would obviously exploit all the kinds of partial order I knew and cared about, I got obsessed enough to solve the speed problem for random data and to minimize the memory burden. The resulting "adaptive, natural, stable" mergesort implemented for Python 2.3 was a major success, but also a major engineering effortthe devil is in the details. There are about 1,200 lines of C code, but unlike the code in the samplesort hybrid, none of these lines are coding for special cases, and about half implement a technical trick allowing the worst-case memory burden to be cut in half. I'm quite proud of it, but the margins of this introduction lack the space for me to explain the details. If you're curious, I wrote a long technical description that you can find in a Python source distribution: Objects/listsort.txt under the main directory (say, Python-2.3.5 or Python-2.4) where you unpacked Python's source distribution archive. In the following list, I provide examples of the partial order Python 2.3's mergesort naturally exploits, where "sorted" means in either forward-sorted or reverse-sorted order: The input is already sorted. The input is mostly sorted but has random elements appended at either end, or both, or inserted in the middle. The input is the concatenation of two or more sorted lists. In fact, the fastest way to merge multiple sorted lists in Python now is to join them into one long list and run list.sort on that. The input is mostly sorted but has some scattered elements that are out of order. This is common, for example, when people manually add new records to a database sorted by name: people aren't good at maintaining strict alphabetic order but are good at getting close. The input has many keys with the same value. For example, when sorting a database of American companies by the stock exchange they're listed on, most will be associated with the NYSE or NASDAQ exchanges. This is exploitable for a curious reason: records with equal keys are already in sorted order, by the definition of "stable"! The algorithm detects that naturally, without code especially looking for equal keys. The input was in sorted order but got dropped on the floor in chunks; the chunks were reassembled in random order, and to fight boredom, some of the chunks were riffle- shuffled together. While that's a silly example, it still results in exploitable partial order and suggests how general the method is. In short, Python 2.3's timsort (well, it has to have some brief name) is stable, robust, and preternaturally fast in many real-life cases: use it any time you can! Recipe 5.1. Sorting a Dictionary Credit: Alex Martelli Problem You want to sort a dictionary. This probably means that you want to sort the keys and then get the values in that same sorted order. Solution The simplest approach is exactly the one expressed by the problem statement: sort the keys, then pick the corresponding values: def sortedDictValues(adict): keys = adict.keys( ) keys.sort( ) return [adict[key] for key in keys] Discussion The concept of sorting applies only to a collection that has an orderin other words, a sequence. A mapping, such as a dictionary, has no order, so it cannot be sorted. And yet, "How do I sort a dictionary?" is a frequent, though literally meaningless, question on the Python lists. More often than not, the question is in fact about sorting some sequence composed of keys and/or values from the dictionary. As for the implementation, while one could think of more sophisticated approaches, it turns out (not unusually, for Python) that the one shown in the solution, the simplest one, is also essentially the fastest one. A further slight increase in speed, about 20%, can be squeezed out in Python 2.3 by replacing the list comprehension with a map call in the return statement at the end of the function. For example: return map(adict.get, keys) Python 2.4, however, is already measurably faster than Python 2.3 with the version in the "Solution" and gains nothing from this further step. Other variants, such as using adict._ _getitem_ _ instead of adict.get, offer no further increase in speed, or they even slow performance down a little, in both Python 2.3 and 2.4. See Also Recipe 5.4 for sorting a dictionary based on its values rather than on its keys. Recipe 5.2. Sorting a List of Strings Case-Insensitively Credit: Kevin Altis, Robin Thomas, Guido van Rossum, Martin V. Lewis, Dave Cross Problem You want to sort a list of strings, ignoring case differences. For example, you want a, although it's lowercase, to sort before B, although the latter is uppercase. By default, however, string comparison is case-sensitive (e.g., all uppercase letters sort before all lowercase ones). Solution The decorate-sort-undecorate (DSU) idiom is simple and fast: def case_insensitive_sort(string_list): auxiliary_list = [(x.lower( ), x) for x in string_list] # decorate auxiliary_list.sort( ) # sort return [x[1] for x in auxiliary_list] # undecorate In Python 2.4, DSU is natively supported, so (assuming the items of string_list are indeed strings, and not, e.g., Unicode objects), you can use the following even shorter and faster approach: def case_insensitive_sort(string_list): return sorted(string_list, key=str.lower) Discussion An obvious alternative to this recipe's Solution is to code a comparison function and pass it to the sort method: def case_insensitive_sort_1(string_list): def compare(a, b): return cmp(a.lower( ), b.lower( )) string_list.sort(compare) However, in this way the lower method gets called twice for every comparison, and the number of comparisons needed to sort a list of n items is typically proportional to n log(n). The DSU idiom builds an auxiliary list, whose items are tuples where each item of the original list is preceded by a "key". The sort then takes place on the key, because Python compares tuples lexicographically (i.e., it compares the tuples' first items first). With DSU, the lower method gets called only n times to sort a list of n strings, which saves enough time to cover the small costs of the first, decorate step and the final, undecorate step, with a big net increase in speed. DSU is also sometimes known, not quite correctly, as the Schwartzian Transform, by somewhat imprecise analogy with a well-known idiom of the Perl language. (If anything, DSU is closer to the Guttman-Rosler Transform, see http://www.sysarch.com/perl/sort_paper.html.) DSU is so important that Python 2.4 supports it directly: you can optionally pass to the sortmethod of a list an argument named key, which is the callable to use on each item of the list to obtain the key for the sort. If you pass such an argument, the sorting internally uses DSU. So, in Python 2.4, string_list.sort(key=str.lower is essentially equivalent to function case_insensitive_sort, except the sort method sorts the list in-place (and returns None) instead of returning a sorted copy and leaving the original list alone. If you want function case_insensitive_sort to sort in-place, by the way, just change its return statement into an assignment to the list's body: string_list[:] = [x[1] for x in auxiliary_list] Vice versa, if, in Python 2.4, you want to get a sorted copy and leave the original list alone, you can use the new built-in function sorted. For example, in Python 2.4: for s in sorted(string_list, key=str.lower): print s prints each string in the list, sorted case-insensitively, without affecting string_list itself. The use of str.lower as the key argument in the Python 2.4 Solution restricts you to specifically sorting strings (not, e.g., Unicode objects). If you know you're sorting a list of Unicode objects, use key=unicode.lower instead. If you need a function that applies just as well to strings and Unicode objects, you can import string and then use key=string.lower; alternatively, you could use key=lambda s: s.lower( ). If you need case-insensitive sorting of lists of strings, you might also need dictionaries and sets using case-insensitive strings as keys, lists behaving case-insensitively regarding such methods as index and count, case-insensitive results from needle in haystack, and so on. If that is the case, then your real underlying need is a subtype of str that behaves case-insensitively in comparison and hashinga clearly better factoring of the issue, compared to implementing many container types and functions to get all of this functionality. To see how to implement such a type, see Recipe 1.24. See Also The Python Frequently Asked Questions http://www.python.org/cgi-bin/faqw.py? req=show&file=faq04.051.htp; Recipe 5.3; Python 2.4 Library Reference about the sorted built- in function and the key argument to sort and sorted; Recipe 1.24. Recipe 5.3. Sorting a List of Objects by an Attribute of the Objects Credit: Yakov Markovitch, Nick Perkins Problem You need to sort a list of objects according to one attribute of each object. Solution The DSU idiom shines, as usual: def sort_by_attr(seq, attr): intermed = [ (getattr(x, attr), i, x) for i, x in enumerate(seq) ] intermed.sort( ) return [ x[-1] for x in intermed ] def sort_by_attr_inplace(lst, attr): lst[:] = sort_by_attr(lst, attr) In Python 2.4, DSU is natively supported, so your code can be even shorter and faster: import operator def sort_by_attr(seq, attr): return sorted(seq, key=operator.attrgetter(attr)) def sort_by_attr_inplace(lst, attr): lst.sort(key=operator.attrgetter(attr)) Discussion Sorting a list of objects by an attribute of each object is best done using the DSU idiom introduced previously in Recipe 5.2. In Python 2.3 and 2.4, DSU is no longer needed, as it used to be, to ensure that a sort is stable (sorting is always stable in Python 2.3 and later), but DSU's speed advantages still shine. Sorting, in the general case and with the best algorithms, is O(n log n) (as is often the case in mathematical formulas, the juxtaposition of terms, here n and log n, indicates that the terms are multiplied). DSU's speed comes from maximally accelerating the O(n log n) part, which dominates sorting time for sequences of substantial length n, by using only Python's native (and maximally fast) comparison. The preliminary decoration step, which prepares an intermediate auxiliary list of tuples, and the successive undecoration step, which extracts the important item from each tuple after the intermediate list is sorted, are only O(n). Therefore any minor inefficiencies in these steps contribute negligible overhead if n is large enough, and reasonably little even for many practical values of n.The O( )-Notation The most useful way to reason about many performance issues is in terms of what is popularly known as big-O analysis and notation (the O stands for "order"). You can find detailed explanations, for example, at http://en.wikipedia.org/wiki/Big_O_notation, but here's a summary. If we consider an algorithm applied to input data of some size N, running time can be described, for large enough values of N (and big inputs are often those for which performance is most critical), as being proportional to some function of N. This is indicated with notations such as O(N) (running time proportional to N: processing twice as much data takes about twice as much time, 10 times as much data, 10 times as much time, and so on; also known as linear time), O(N squared) (running time proportional to the square of N: processing twice as much data takes about four times as much time, 10 times as much data, 100 times as much time; also known as quadratic time), and so on. Another case you will see often is O(N log N), which is faster than O(N squared) but not as fast as O(N). The constant of proportionality is often ignored (at least in theoretical analysis) because it depends on such issues as the clock rate of your computer, not just on the algorithm. If you buy a machine that's twice as fast as your old one, everything will run in half the time, but that will not change any of the comparisons between alternative algorithms. This recipe puts index i, in each tuple that is an item of list intermed, ahead of the corresponding x (where x is the i-th item in seq). This placement ensures that two items of seq will never be compared directly, even if they have the same value for the attribute named attr. Even in that case, their indices will still differ, and thus Python's lexicographic comparison of the tuples will never get all the way to comparing the tuples' last items (the original items from seq). Avoiding object comparisons may save us from performing extremely slow operations, or even from attempting forbidden ones. For example, we could sort a list of complex numbers by their real attribute: we would get an exception if we ever tried to compare two complex numbers directly, because no ordering is defined on complex numbers. But thanks to the precaution described in this paragraph, such an event can never occur, and the sorting will therefore proceed correctly. As mentioned earlier in Recipe 5.2, Python 2.4 supports DSU directly. You can pass an optional keyword-argument key, to sort, which is the callable to use on each item to get the sort key. Standard library module operator has two new functions, attrgetter and itemgetter, that exist specifically to return callables suitable for this purpose. In Python 2.4, the ideal solution to this problem therefore becomes: import operator seq.sort(key=operator.attrgetter(attr)) This snippet performs the sort in-place, which helps make it blazingly faston my machine, three times faster than the Python 2.3 function shown first in this recipe. If you need a sorted copy, without disturbing seq, you can get it using Python 2.4's new built-in function sorted: sorted_copy = sorted(seq, key=operator.attrgetter(attr)) While not quite as fast as an in-place sort, this latest snippet is still over 2.5 times faster than the function shown first in this recipe. Python 2.4 also guarantees that, when you pass the optionalkey named argument, list items will never be accidentally compared directly, so you need not take any special safeguards. Moreover, stability is also guaranteed. See Also Recipe 5.2; Python 2.4's Library Reference docs about the sorted built-in function, operator module's attrgetter and itemgetter functions, and the key argument to .sort and sorted. Recipe 5.4. Sorting Keys or Indices Basedon the Corresponding Values Credit: John Jensen, Fred Bremmer, Nick Coghlan Problem You need to count the occurrences of various items and present the items in order of their number of occurrencesfor example, to produce a histogram. Solution A histogram, apart from graphical issues, is based on counting the occurrences of items (easy to do with a Python list or dictionary) and then sorting the keys or indices in an order based on corresponding values. Here is a subclass of dict that adds two methods for the purpose: class hist(dict): def add(self, item, increment=1): ''' add 'increment' to the entry for 'item' ''' self[item] = increment + self.get(item, 0) def counts(self, reverse=False): ''' return list of keys sorted by corresponding values ''' aux = [ (self[k], k) for k in self ] aux.sort( ) if reverse: aux.reverse( ) return [k for v, k in aux] If the items you're counting are best modeled by small integers in a compact range, so that you want to keep item counts in a list, the solution is quite similar: class hist1(list): def _ _init_ _(self, n): ''' initialize this list to count occurrences of n distinct items ''' list._ _init_ _(self, n*[0]) def add(self, item, increment=1): ''' add 'increment' to the entry for 'item' ''' self[item] += increment def counts(self, reverse=False): ''' return list of indices sorted by corresponding values ''' aux = [ (v, k) for k, v in enumerate(self) ] aux.sort( ) if reverse: aux.reverse( ) return [k for v, k in aux] Discussion The add method of hist embodies the normal Python idiom for counting occurrences of arbitrary (but hashable) items, using a dict to hold the counts. In class hist1, based on a list, we take a different approach, initializing all counts to 0 in _ _init_ _, so the add method is even simpler.The counts methods produce the lists of keys, or indices, sorted in the order given by the corresponding values. The problem is very similar in both classes, hist and hist1; therefore, the solutions are also almost identical, using in each case the DSU approach already shown in Recipe 5.2 and Recipe 5.3. If we need both classes in our program, the similarity is so close that we should surely factor out the commonalities into a single auxiliary function _sorted_keys: def _sorted_keys(container, keys, reverse): ''' return list of 'keys' sorted by corresponding values in 'container' ''' aux = [ (container[k], k) for k in keys ] aux.sort( ) if reverse: aux.reverse( ) return [k for v, k in aux] and then implement the counts methods of each class as thin wrappers over this _sorted_keys function: class hist(dict): ... def counts(self, reverse=False): return _sorted_keys(self, self, reverse) class hist1(list): ... def counts(self, reverse=False): return _sorted_keys(self, xrange(len(self)), reverse) DSU is so important that in Python 2.4, as shown previously in Recipe 5.2 and Recipe 5.3, the sort method of lists and the new built-in function sorted offer a fast, intrinsic implementation of it. Therefore, in Python 2.4, function _sorted_keys can become much simpler and faster: def _sorted_keys(container, keys, reverse): return sorted(keys, key=container._ _getitem_ _, reverse=reverse)` The bound-method container._ _getitem_ _ performs exactly the same operation as the indexing container[k] in the Python 2.3 implementation, but it's a callable to call on each k of the sequence that we're sorting, namely keysexactly the right kind of value to pass as the key keyword argument to the sorted built-in function. Python 2.4 also affords an easy, direct way to get a list of a dictionary's items sorted by value: from operator import itemgetter def dict_items_sorted_by_value(d, reverse=False): return sorted(d.iteritems( ), key=itemgetter(1), reverse=reverse) The operator.itemgetter higher-order function, also new in Python 2.4, is a handy way to supply the key argument when you want to sort a container whose items are subcontainers, keying on a certain item of each subcontainer. This is exactly the case here, since a dictionary's items are a sequence of pairs (two-item tuples), and we want to sort the sequence keying on the second item of each tuple. Getting back to this recipe's main theme, here is a usage example for the class hist shown in this recipe's Solution: sentence = ''' Hello there this is a test. Hello there this was a test, but now it is not. ''' words = sentence.split( ) c = hist( ) for word in words: c.add(word)print "Ascending count:" print c.counts( ) print "Descending count:" print c.counts(reverse=True) This code snippet produces the following output: Ascending count: [(1, 'but'), (1, 'it'), (1, 'not.'), (1, 'now'), (1, 'test,'), (1, 'test.'), (1, 'was'), (2, 'Hello'), (2, 'a'), (2, 'is'), (2, 'there'), (2, 'this')] Descending count: [(2, 'this'), (2, 'there'), (2, 'is'), (2, 'a'), (2, 'Hello'), (1, 'was'), (1, 'test.'), (1, 'test,'), (1, 'now'), (1, 'not.'), (1, 'it'), (1, 'but')] See Also Recipe "Special Method Names" in the Language Reference and the OOP chapter in Python in a Nutshell, about special method _ _getitem_ _; Library Reference docs for Python 2.4 sorted built-in and the key= argument to sort and sorted. Recipe 5.5. Sorting Strings with Embedded Numbers Credit: Sébastien Keim, Chui Tey, Alex Martelli Problem You need to sort a list of strings that contain substrings of digits (e.g., a list of postal addresses) in an order that looks good. For example, 'foo2.txt' should come before 'foo10.txt'. However, Python's default string comparison is alphabetical, so, by default, 'foo10.txt' instead comes before 'foo2.txt'. Solution You need to split each string into sequences of digits and nondigits, and transform each sequence of digits into a number. This gives you a list that is just the right comparison key for the sort you want, and you can then use DSU for the sort itselfthat is, code two functions, shorter than this description: import re re_digits = re.compile(r'(\d+)') def embedded_numbers(s): pieces = re_digits.split(s) # split into digits/nondigits pieces[1::2] = map(int, pieces[1::2]) # turn digits into numbers return pieces def sort_strings_with_embedded_numbers(alist): aux = [ (embedded_numbers(s), s) for s in alist ] aux.sort( ) return [ s for _ _, s in aux ] # convention: _ _ means "ignore" In Python 2.4, use the native support for DSU, with the same function embedded_numbers to get the sort key: def sort_strings_with_embedded_numbers(alist): return sorted(alist, key=embedded_numbers) Discussion Say you have an unsorted list of filenames, such as: files = 'file3.txt file11.txt file7.txt file4.txt file15.txt'.split( ) If you just sort and print this list, for example in Python 2.4 with print ' '.join(sorted(files)), your output looks like file11.txt file15.txt file3.txt file4.txt file7.txt, since, by default, strings are sorted alphabetically (to use a fancier word, the sort order is described as lexicographical). Python cannot just guess that you mean to treat in a different way those substrings that happen to be made of digits; you have to tell Python precisely what you want, and this recipe shows how. Using this recipe, you can get a nicer-looking result:print ' '.join(sort_strings_with_embedded_numbers(files)) The output is now file3.txt file4.txt file7.txt file11.txt file15.txt, which is probably just what you want in this case. The implementation relies on the DSU idiom. We need to code DSU explicitly if we want to support Python 2.3, while if our code is Python 2.4-only, we just rely on the native implementation of DSU. We do so by passing an argument named key (a function to be called on each item to get the right comparison key for the sort) to the new built-in function sorted. Function embedded_numbers in the recipe is how we get the right comparison key for each item: a list alternating substrings of nondigits, and the int obtained from each substring of digits. re_digits.split(s) gives us a list of alternating substrings of nondigits and digits (with the substrings of digits at odd-numbered indices); then, we use built-in functions map and int (and extended-form slices that get and set all items at odd-numbered indices) to turn sequences of digits into integers. Lexicographical comparison on this list of mixed types now produces just the right result. See Also Library Reference and Python in a Nutshell docs about extended slicing and about module re; Python 2.4 Library Reference about the sorted built-in function and the key argument to sort and sorted; Recipe 5.3; Recipe 5.2. Recipe 5.6. Processing All of a List's Items in Random Order Credit: Iuri Wickert, Duncan Grisby, T. Warner, Steve Holden, Alex Martelli Problem You need to process, in random order, all of the items of a long list. Solution As usual in Python, the best approach is the simplest one. If we are allowed to change the order of items in the input list, then the following function is simplest and fastest: def process_all_in_random_order(data, process): # first, put the whole list into random order random.shuffle(data) # next, just walk over the list linearly for elem in data: process(elem) If we must preserve the input list intact, or if the input data may be some iterable that is not a list, just insert as the first statement of the function the assignment data = list(data). Discussion While it's a common mistake to be overly concerned with speed, don't make the opposite mistake of ignoring the different performances of various algorithms. Suppose we must process all of the items in a long list in random order, without repetition (assume that we're allowed to mangle or destroy the input list). The first idea to suggest itself might be to repeatedly pick an item at random (with function random.choice), removing each picked item from the list to avoid future repetitions: import random def process_random_removing(data, process): while data: elem = random.choice(data) data.remove(elem) process(elem) However, this function is painfully slow, even for input lists of just a few hundred elements. Each call to data.remove must linearly search through the list to find the element to delete. Since the cost of each of n steps is O(n), the whole process is O(n2)time proportional to the square of the length of the list (and with a large multiplicative constant, too). Minor improvements to this first idea could focus on obtaining random indices, using the pop method of the list to get and remove an item at the same time, low-level fiddling with indices to avoid the costly removal in favor of swapping the picked item with the last yet-unpicked one towards the end, or using dictionaries or sets instead of lists. This latest idea might be based ona hope of using a dict's popitem method (or the equivalent method pop of class sets.Set and Python 2.4's built-in type set), which may look like it's designed exactly to pick and remove a random item, but, beware! dict.popitem is documented to return and remove an arbitrary item of the dictionary, and that's a far cry from a random item. Check it out: >>> d=dict(enumerate('ciao')) >>> while d: print d.popitem( ) It may surprise you, but in most Python implementations this snippet will print d's items in a far from random order, typically (0,'c') then (1,'i') and so forth. In short, if you need pseudo- random behavior in Python, you need standard library module randompopitem is not an alternative. If you thought about using a dictionary rather than a list, you are definitely on your way to "thinking Pythonically", even though it turns out that dictionaries wouldn't provide a substantial performance boost for this specific problem. However, an approach that is even more Pythonic than choosing the right data structure is best summarized as: let the standard library do it!. The Python Standard Library is large, rich, and chock full of useful, robust, fast functions and classes for a wide variety of tasks. In this case, the key intuition is realizing that, to walk over a sequence in a random order, the simplest approach is to first put that sequence into random order (known as shuffling the sequence, an analogy with shuffling a deck of cards) and then walk over the shuffled sequence linearly. Function random.shuffle performs the shuffling, and the function shown in this recipe's Solution just uses it. Performance should always be measured, never guessed at, and that's what standard library module timeit is for. Using a null process function and a list of length 1,000 as data, process_all_in_random_order is almost 10 times faster than process_random_removing; with a list of length 2,000, the performance ratio grows to almost 20. While an improvement of, say, 25%, or even a constant factor of 2, usually can be neglected without really affecting the performance of your program as a whole, the same does not apply to an algorithm that is 10 or 20 times as slow as it could be. Such terrible performance is likely to make that program fragment a bottleneck, all by itself. Moreover, this risk increases when we're talking about O(n2) versus O(n) behavior: with such differences in big-O behavior, the performance ratio between bad and good algorithms keeps increasing without bounds as the size of the input data grows. See Also The documentation for the random and timeit modules in the Library Reference and Python in a Nutshell. Recipe 5.7. Keeping a Sequence Ordered as Items Are Added Credit: John Nielsen Problem You want to maintain a sequence, to which items are added, in a sorted state, so that at any time, you can easily examine or remove the smallest item currently present in the sequence. Solution Say you start with an unordered list, such as: the_list = [903, 10, 35, 69, 933, 485, 519, 379, 102, 402, 883, 1] You could call the_list.sort( ) to make the list sorted and then result=the_list.pop(0) to get and remove the smallest item. But then, every time you add an item (say with the_list.append(0)), you need to call the_list.sort( ) again to keep the list sorted. Alternatively, you can use the heapq module of the Python Standard Library: import heapq heapq.heapify(the_list) Now the list is not necessarily fully sorted, but it does satisfy the heap property (meaning if all indices involved are valid, the_list[i]<=the_list[2*i+1] and the_list[i]<=the_list[2*i+2])so, in particular, the_list[0] is the smallest item. To keep the heap property valid, use result=heapq.heappop(the_list) to get and remove the smallest item and heapq.heappush(the_list, newitem) to add a new item. When you need to do bothadd a new item while getting and removing the previously smallest itemyou can use result=heapq.heapreplace(the_list, newitem). Discussion When you need to retrieve data in an ordered way (at each retrieval getting the smallest item among those you currently have at hand), you can pay the runtime cost for the sorting when you retrieve the data, or you can pay for it when you add the data. One approach is to collect your data into a list and sort the list. Now it's easy to get your data in order, smallest to largest. However, you have to keep calling sort each time you add new data during the retrieval, to make sure you can later keep retrieving from the smallest current item after each addition. The method sort of Python lists is implemented with a little-known algorithm called Natural Mergesort, which minimizes the runtime cost of this approach. Yet the approach can still be burdensome: each addition (and sorting) and each retrieval (and removal, via pop) takes time proportional to the number of current items in the list (O(N), in common parlance). An alternative approach is to use a data organization known as a heap, a type of binary tree implemented compactly, yet ensuring that each "parent" is always less than its "children". Thebest way to maintain a heap in Python is to use a list and have it managed by the heapq library module, as shown in this recipe's Solution. The list does not get fully sorted, yet you can be sure that, whenever you heappop an item from the list, you always get the lowest item currently present, and all others will be adjusted to ensure the heap property is still valid. Each addition with heappush, and each removal with heappop, takes a short time proportional to the logarithm of the current length of the list (O(log N), in common parlance). You pay as you go, a little at a time (and not too much in total, either.) A good occasion to use this heap approach, for example, is when you have a long-running queue with new data periodically arriving, and you always want to be able to get the most important item off the queue without having to constantly re-sort your data or perform full searches. This concept is known as a priority queue, and a heap is an excellent way to implement it. Note that, intrinsically, the heapq module supplies you with the smallest item at each heappop, so make sure to arrange the way you encode your items' priority values to reflect this. For example, say that you receive incoming items each accompanied by a cost, and the most important item at any time is the one with the highest cost that is currently on the queue; moreover, among items of equal cost, the most important one is the one that arrived earliest. Here's a way to build a "priority queue" class respecting these specs and based on functions of module heapq: class prioq(object): def _ _init_ _(self): self.q = [ ] self.i = 0 def push(self, item, cost): heapq.heappush(self.q, (-cost, self.i, item)) self.i += 1 def pop(self): return heapq.heappop(self.q)[-1] The main idea in this snippet is to push on the heap tuples whose first item is the cost with changed sign, so that higher costs result in smaller tuples (by Python's natural comparison); right after the cost, we put a progressive index, so that, among items with equal cost, the one arriving earliest will be in a smaller tuple. In Python 2.4, module heapq has been reimplemented and optimized; see Recipe 5.8 for more information about heapq. See Also Docs for module heapq in the Library Reference and Python in a Nutshell; heapq.py in the Python sources contains a very interesting discussion of heaps; Recipe 5.8 for more information about heapq; Recipe 19.14 for merging sorted sequences using heapq. Recipe 5.8. Getting the First Few Smallest Items of a Sequence Credit: Matteo Dell'Amico, Raymond Hettinger, George Yoshida, Daniel Harding Problem You need to get just a few of the smallest items from a sequence. You could sort the sequence and just use seq[:n], but is there any way you can do better? Solution Perhaps you can do better, if n, the number of items you need, is small compared to n, the sequence's length. sort is very fast, but it still takes O(n log n) time, while we can get the first n smallest elements in time O(n) if n is small. Here is a simple and practical generator for this purpose, which works equally well in Python 2.3 and 2.4: import heapq def isorted(data): data = list(data) heapq.heapify(data) while data: yield heapq.heappop(data) In Python 2.4 only, you can use an even simpler and faster way to get the smallest n items of data when you know n in advance: import heapq def smallest(n, data): return heapq.nsmallest(n, data) Discussion data can be any bounded iterable; the recipe's function isorted starts by calling list on it to ensure that. You can remove the statement data = list(data) if all the following conditions hold: you know that data is a list to start with, you don't mind the fact that the generator reorders data's items, and you want to remove items from data as you fetch them. As shown previously in Recipe 5.7, the Python Standard Library contains module heapq, which supports the data structures known as heaps. Generator isorted in this recipe's Solution relies on making a heap at the start (via heap.heapify) and then yielding and removing the heap's smallest remaining item at each step (via heap.heappop). In Python 2.4, module heapq has also grown two new functions. heapq.nlargest(n, data) returns a list of the n largest items of data; heapq.nsmallest(n, data) returns a list of the n smallest items. These functions do not require that data satisfy the heap condition; indeed, they do not even require data to be a listany bounded iterable whose items are comparable will do. Function smallest in this recipe's Solution just lets heapq.smallest do all the work.To judge speed, we must always measure itguessing about relative speeds of different pieces of code is a mug's game. So, how does isorted's performance compare with Python 2.4's built-in function sorted, when we're only looping on the first few (smallest) items? To help measure timing, I wrote a top10 function that can use either approach, and I also made sure I had a sorted function even in Python 2.3, where it's not built in: try: sorted except: def sorted(data): data = list(data) data.sort( ) return data import itertools def top10(data, howtosort): return list(itertools.islice(howtosort(data), 10)) On my machine running Python 2.4 on thoroughly shuffled lists of 1,000 integers, top10 takes about 260 microseconds with isorted, while it takes about 850 microseconds with the built-in sorted. However, Python 2.3 is much slower and gives vastly different results: about 12 milliseconds with isorted, about 2.7 milliseconds with sorted. In other words, Python 2.3 is 3 times slower than Python 2.4 for sorted, but it's 50 times slower for isorted. Lesson to retain: whenever you optimize, measure. You shouldn't choose optimizations based on first principles, since the performance numbers can vary so widely, even between vastly compatible "point releases". A secondary point can be made: if you care about performance, move to Python 2.4 as soon as you can. Python 2.4 has been vastly optimized and accelerated over Python 2.3, particularly in areas related to searching and sorting. If you know that your code need only support Python 2.4, then, as this recipe's Solution indicates, using heapq's new function nsmallest is faster, as well as simpler, than doing your own coding. To implement top10 in Python 2.4, for example, you just need: import heapq def top10(data): return heapq.nsmallest(10, data) This version takes about half the time of the previously shown isorted-based top10, when called on the same thoroughly shuffled lists of 1,000 integers. See Also Library Reference and Python in a Nutshell docs about method sort of type list, and about modules heapq and timeit; Chapter 19 for more about iteration in Python; Python in a Nutshell's chapter on optimization; heapq.py in the Python sources contains a very interesting discussion of heaps; Recipe 5.7 for more information about heapq. Recipe 5.9. Looking for Items in a Sorted Sequence Credit: Noah Spurrier Problem You need to look for a lot of items in a sequence. Solution If list L is sorted, module bisect from the Python Standard Library makes it easy to check if some item x is present in L: import bisect x_insert_point = bisect.bisect_right(L, x) x_is_present = L[x_insert_point-1:x_insert_point] == [x] Discussion Looking for an item x in a list L is very easy in Python: to check whether the item is there at all, if x in L; to find out where exactly it is, L.index(x). However, if L has length n, these operations take time proportional to nessentially, they just loop over the list's items, checking each for equality to x. If L is sorted, we can do better. The classic algorithm to look for an item in a sorted sequence is known as binary search, because at each step it roughly halves the range it's still searching onit generally takes about log2n steps. It's worth considering when you're going to look for items many times, so you can amortize the cost of sorting over many searches. Once you've decided to use binary search for x in L, after calling L.sort( ), module bisect from the Python Standard Library makes the job easy. Specifically, we need function bisect.bisect_right, which returns the index where an item should be inserted, to keep the sorted list sorted, but doesn't alter the list; moreover, if the item already appears in the list, bisect_right returns an index that's just to the right of any items with the same value. So, after getting this "insert point" by calling bisect.bisect_right(L, x), we need only to check the list immediately before the insert point, to see if an item equal to x is already there. The way we compute x_is_present in the "Solution" may not be immediately obvious. If we know that L is not empty, we can use a simpler and more obvious approach: x_is_present = L[x_insert_point-1] == x However, the indexing in this simpler approach raises an exception when L is empty. When the slice boundaries are invalid, slicing is less "strict" than indexing, since it just produces an empty slice without raising any exception. In general, somelist[i:i+1] is the same one-item list as [somelist[i]] when i is a valid index in somelist: it's an empty list [ ] when the indexing would raise IndexError. The computation of x_is_present in the recipe exploits this important property to avoid having to deal with exceptions and handle empty and nonempty cases for L in one uniform way. An alternative approach is:x_is_present = L and L[x_insert_point-1] == x This alternative approach exploits and's short-circuiting behavior to guard the indexing, instead of using slicing. An auxiliary dict, as shown in Recipe 5.12, is also a possibility as long as items are hashable (meaning that items can be used as keys into a dict). However, the approach in this recipe, based on a sorted list, may be the only useful one when the items are comparable (otherwise, the list could not be sorted) but not hashable (so a dict can't have those items as its keys). When the list is already sorted, and the number of items you need to look up in it is not extremely large, it may in any case be faster to use bisect than to build an auxiliary dictionary, since the investment of time in the latter operation might not be fully amortized. This is particularly likely in Python 2.4, since bisect has been optimized very effectively and is much faster than it was in Python 2.3. On my machine, for example, bisect.bisect_right for an item in the middle of a list of 10,000 integers is about four times faster in Python 2.4 than it was in Python 2.3. See Also Documentation for the bisect module in the Library Reference and Python in a Nutshell; Recipe 5.12. Recipe 5.10. Selecting the nth Smallest Element of a Sequence Credit: Raymond Hettinger, David Eppstein, Shane Holloway, Chris Perkins Problem You need to get from a sequence the nth item in rank order (e.g., the middle item, known as the median). If the sequence was sorted, you would just use seq[n]. But the sequence isn't sorted, and you wonder if you can do better than just sorting it first. Solution Perhaps you can do better, if the sequence is big, has been shuffled enough, and comparisons between its items are costly. Sort is very fast, but in the end (when applied to a thoroughly shuffled sequence of length n) it always takes O(n log n) time, while there exist algorithms that can be used to get the nth smallest element in time O(n). Here is a function with a solid implementation of such an algorithm: import random def select(data, n): " Find the nth rank ordered element (the least value has rank 0). " # make a new list, deal with <0 indices, check for valid index data = list(data) if n<0: n += len(data) if not 0 <= n < len(data): raise ValueError, "can't get rank %d out of %d" % (n, len(data)) # main loop, quicksort-like but with no need for recursion while True: pivot = random.choice(data) pcount = 0 under, over = [ ], [ ] uappend, oappend = under.append, over.append for elem in data: if elem < pivot: uappend(elem) elif elem > pivot: oappend(elem) else: pcount += 1 numunder = len(under) if n < numunder: data = under elif n < numunder + pcount: return pivot else: data = over n -= numunder + pcountDiscussion This recipe is meant for cases in which repetitions count. For example, the median of the list [1, 1, 1, 2, 3] is 1 because that is the third one of the five items in rank order. If, for some strange reason, you want to discount duplications, you need to reduce the list to its unique items first (e.g., by applying the Recipe 18.1), after which you may want to come back to this recipe. Input argument data can be any bounded iterable; the recipe starts by calling list on it to ensure that. The algorithm then loops, implementing at each leg a few key ideas: randomly choosing a pivot element; slicing up the list into two parts, made up of the items that are "under" and "over" the pivot respectively; continuing work for the next leg on just one of the two parts, since we can tell which one of them the nth element will be in, and the other part can safely be ignored. The ideas are very close to that in the classic algorithm known as quicksort (except that quicksort cannot ignore either part, and thus must use recursion, or recursion-removal techniques such as keeping an explicit stack, to make sure it deals with both parts). The random choice of pivot makes the algorithm robust against unfavorable data orderings (the kind that wreak havoc with naive quicksort); this implementation decision costs about log2N calls to random.choice. Another implementation issue worth pointing out is that the recipe counts the number of occurrences of the pivot: this precaution ensures good performance even in the anomalous case where data contains a high number of repetitions of identical values. Extracting the bound methods .append of lists under and over as local variables uappend and oappend may look like a pointless, if tiny, complication, but it is, in fact, a very important optimization technique in Python. To keep the compiler simple, straightforward, unsurprising, and robust, Python does not hoist constant computations out of loops, nor does it "cache" the results of method lookup. If you call under.append and over.append in the inner loop, you pay the cost of lookup each and every time. If you want something hoisted, hoist it yourself. When you're considering an optimization, you should always measure the code's performance with and without that optimization, to check that the optimization does indeed make an important difference. According to my measurements, removing this single optimization slows performance down by about 50% for the typical task of picking the 5000th item of range(10000). Considering the tiny amount of complication involved, a difference of 50% is well worth it. A natural idea for optimization, which just didn't make the grade once carefully measured, is to call cmp(elem, pivot) in the loop body, rather than making separate tests for elem < pivot and elem > pivot. Unfortunately, measurement shows that cmp doesn't speed things up; in fact, it slows them down, at least when the items of data are of elementary types such as numbers and strings. So, how does select's performance compare with the simpler alternative of: def selsor(data, n): data = list(data) data.sort( ) return data[n] On thoroughly shuffled lists of 3,001 integers on my machine, this recipe's select takes about 16 milliseconds to find the median, while selsor takes about 13 milliseconds; considering that sort could take advantage of any partial sortedness in the data, for this kind of length, and on elementary data whose comparisons are fast, it's not to your advantage to use select. For a length of 30,001, performance becomes very close between the two approachesaround 170 milliseconds either way. When you push the length all the way to 300,001, select provides an advantage, finding the median in about 2.2 seconds, while selsor takes about 2.5. The break-even point will be smaller if the items in the sequence have costly comparison methods, since the key difference between the two approaches is in the number of comparisons performedselect takes O(n), selsor takes O(n log n). For example, say we need to compare instances of a class designed for somewhat costly comparisons (simulating four-dimensionalpoints that will often coincide on the first few dimensions): class X(object): def _ _init_ _(self): self.a = self.b = self.c = 23.51 self.d = random.random( ) def _dats(self): return self.a, self.b, self.c, self.d def _ _cmp_ _(self, oth): return cmp(self._dats, oth._dats) Here, select already becomes faster than selsor when what we're computing is the median of vectors of 201 such instances. In other words, although select has more general overhead, when compared to the wondrously efficient coding of lists' sort method, nevertheless, if n is large enough and each comparison is costly enough, select is still well worth considering. See Also Library Reference and Python in a Nutshell docs about method sort of type list, and about module random. Recipe 5.11. Showing off quicksort in Three Lines Credit: Nathaniel Gray, Raymond Hettinger, Christophe Delord, Jeremy Zucker Problem You need to show that Python's support for the functional programming paradigm is better than it might seem at first sight. Solution Functional programming languages, of which Haskell is a great example, are splendid animals, but Python can hold its own in such company: def qsort(L): if len(L) <= 1: return L return qsort([lt for lt in L[1:] if lt < L[0]]) + L[0:1] + \ qsort([ge for ge in L[1:] if ge >= L[0]]) In my humble opinion, this code is almost as pretty as the Haskell version from http://www.haskell.org: qsort [ ] = [ ] qsort (x:xs) = qsort elts_lt_x ++ [x] ++ qsort elts_greq_x where elts_lt_x = [y | y <- xs, y < x] elts_greq_x = [y | y <- xs, y >= x] Here's a test function for the Python version: def qs_test(length): import random joe = range(length) random.shuffle(joe) qsJoe = qsort(joe) for i in range(len(qsJoe)): assert qsJoe[i] == i, 'qsort is broken at %d!' %i Discussion This rather naive implementation of quicksort illustrates the expressive power of list comprehensions. Do not use this approach in real code! Python lists have an in-place sort method that is much faster and should always be preferred; in Python 2.4, the new built-in function sorted accepts any finite sequence and returns a new sorted list with the sequence's items. The only proper use of this recipe is for impressing friends, particularly ones who (quite understandably) are enthusiastic about functional programming, and particularly about the Haskell language. I cooked up this function after finding the wonderful Haskell quicksort (which I've reproduced inthe "Solution") at http://www.haskell.org/aboutHaskell.html. After marveling at the elegance of this code for a while, I realized that list comprehensions made the same approach possible in Python. Not for nothing did we steal list comprehensions right out of Haskell, just Pythonizing them a bit by using keywords rather than punctuation! Both implementations pivot on the first element of the list and thus have worst-case O(n) performance for the very common case of sorting an already sorted list. You would never want to do so in production code! Because this recipe is just a propaganda piece, though, it doesn't really matter. You can write a less compact version with similar architecture in order to use named local variables and functions for enhanced clarity: def qsort(L): if not L: return L pivot = L[0] def lt(x): return x=pivot return qsort(filter(lt, L[1:]))+[pivot]+qsort(filter(ge, L[1:])) Once you start going this route, you can easily move to a slightly less naive version, using random pivot selection to make worst-case performance less likely and counting pivots to handle degenerate case with many equal elements: import random def qsort(L): if not L: return L pivot = random.choice(L) def lt(x): return xpivot return qsort(filter(lt, L))+[pivot]*L.count(pivot)+qsort(filter(gt, L)) Despite the enhancements, they are meant essentially for fun and demonstration purposes. Production-quality sorting code is quite another thing: these little jewels, no matter how much we dwell on them, will never match the performance and solidity of Python's own built-in sorting approaches. Rather than going for clarity and robustness, we can move in the opposite direction to make this last point most obvious, showing off the obscurity and compactness that one can get with Python's lambda: q=lambda x:(lambda o=lambda s:[i for i in x if cmp(i,x[0])==s]: len(x)>1 and q(o(-1))+o(0)+q(o(1)) or x)( ) At least, with this beauty (a single logical line, although it needs to be split into two physical lines due to its length), it should be absolutely obvious that this approach is not meant for real-world use. The equivalent, using more readable def statements rather than opaque lambdas, would still be pretty obscure: def q(x): def o(s): return [i for i in x if cmp(i,x[0])==s] return len(x)>1 and q(o(-1))+o(0)+q(o(1)) or x but a little more clarity (and sanity) can be recovered by opening up the pithy len(x)>1 and . . . or x into an if/else statement and introducing sensible local names again: def q(x): if len(x)>1: lt = [i for i in x if cmp(i,x[0]) == -1 ] eq = [i for i in x if cmp(i,x[0]) == 0 ] gt = [i for i in x if cmp(i,x[0]) == 1 ] return q(lt) + eq + q(gt) else: return x Fortunately, in the real world, Pythonistas are much too sensible to write convoluted, lambda- filled horrors such as this. In fact, many (though admittedly not all) of us feel enough aversion to lambda itself (partly from having seen it abused this way) that we go out of our way to use readable def statements instead. As a result, the ability to decode such "bursts of line noise" is not a necessary survival skill in the Python world, as it might be for other languages. Any language feature can be abused by programmers trying to be "clever" . . . as a result, some Pythonistas (though a minority) feel a similar aversion to features such as list comprehensions (since it's possible to cram too many things into a list comprehension, where a plain for loop would be clearer) or to the short-circuiting behavior of operators and/or (since they can be abused to write obscure, terse expressions where a plain if statement would be clearer). See Also The Haskell web site, http://www.haskell.org. Recipe 5.12. Performing Frequent Membership Tests on a Sequence Credit: Alex Martelli Problem You need to perform frequent tests for membership in a sequence. The O(n) behavior of repeated in operators hurts performance, but you can't switch to using just a dictionary or set instead of the sequence, because you also need to keep the sequence's order. Solution Say you need to append items to a list only if they're not already in the list. One sound approach to this task is the following function: def addUnique(baseList, otherList): auxDict = dict.fromkeys(baseList) for item in otherList: if item not in auxDict: baseList.append(item) auxDict[item] = None If your code has to run only under Python 2.4, you can get exactly the same effect with an auxiliary set rather than an auxiliary dictionary. Discussion A simple (naive?) approach to this recipe's task looks good: def addUnique_simple(baseList, otherList): for item in otherList: if item not in baseList: baseList.append(item) and it may be sort of OK, if the lists are very small. However, the simple approach can be quite slow if the lists are not small. When you check if item not in baseList, Python can implement the in operator in only one way: an internal loop over the elements of baseList, ending with a result of true as soon as an element compares equal to item, with a result of False if the loop terminates without having found any equality. On average, executing the in-operator takes time proportional to len(baseList). addUnique_simple executes the in-operator len(otherList) times, so, in all, it takes time proportional to the product of the lengths of the two lists. In the addUnique function shown in the "Solution", we first build the auxiliary dictionary auxDict, a step that takes time proportional to len(baseList). Then, the in-operator inside the loop checks for membership in a dicta step that makes all the difference because checking formembership in a dict takes roughly constant time, independent of the number of items in the dict! So, the for loop takes time proportional to len(otherList), and the entire function takes time proportional to the sum of the lengths of the two lists. The analysis of the running times should in fact go quite a bit deeper, because the length of baseList is not constant in addUnique_simple; baseList grows each time an item is processed that was not already there. But the gist of the (surprisingly complicated) analysis is not very different from what this simplified version indicates. We can check this by measuring. When each list holds 10 integers, with an overlap of 50%, the simple version is about 30% slower than the one shown in the "Solution", the kind of slowdown that can normally be ignored. But with lists of 100 integers each, again with 50% overlap, the simple version is twelve times slower than the one shown in the "Solution"a level of slowdown that can never be ignored, and it only gets worse if the lists get really substantial. Sometimes, you could obtain even better overall performance for your program by permanently placing the auxiliary dict alongside the sequence, encapsulating both into one object. However, in this case, you must maintain the dict as the sequence gets modified, to ensure it stays in sync with the sequence's current membership. This maintenance task is not trivial, and it can be architected in many different ways. Here is one such way, which does the syncing "just in time," rebuilding the auxiliary dict when a membership test is required and the dictionary is possibly out of sync with the list's contents. Since it costs very little, the following class optimizes the index method, as well as membership tests: class list_with_aux_dict(list): def _ _init_ _(self, iterable=( )): list._ _init_ _(self, iterable) self._dict_ok = False def _rebuild_dict(self): self._dict = { } for i, item in enumerate(self): if item not in self._dict: self._dict[item] = i self._dict_ok = True def _ _contains_ _(self, item): if not self._dict_ok: self._rebuild_dict( ) return item in self._dict def index(self, item): if not self._dict_ok: self._rebuild_dict( ) try: return self._dict[item] except KeyError: raise ValueError def _wrapMutatorMethod(methname): _method = getattr(list, methname) def wrapper(self, *args): # Reset 'dictionary OK' flag, then delegate to the real mutator method self._dict_ok = False return _method(self, *args) # in Python 2.4, only: wrapper._ _name_ _ = _method._ _name_ _ setattr(list_with_aux_dict, methname, wrapper) for meth in 'setitem delitem setslice delslice iadd'.split( ): _wrapMutatorMethod('_ _%s_ _' % meth) for meth in 'append insert pop remove extend'.split( ): _wrapMutatorMethod(meth) del _wrapMethod # remove auxiliary function, not needed any more The list_with_aux_dict class extends list and delegates to it every method, except _ _contains_ _ and index. Every method that can modify list membership is wrapped in a closure that resets a flag asserting that the auxiliary dictionary is OK. Python's in-operator calls the _ _contains_ _ method. list_with_aux_dict's _ _contains_ _ method rebuilds the auxiliarydictionary, unless the flag is set (when the flag is set, rebuilding is unnecessary); the index method works the same way. Instead of building and installing wrapping closures for all the mutating methods of the list into the list_with_aux_dict class with a helper function, as the recipe does, we could write all the def statements for the wrapper methods in the body of list_with_aux_dict. However, the code for the class as presented has the important advantage of minimizing boilerplate (repetitious plumbing code that is boring and voluminous, and thus a likely home for bugs). Python's strengths at introspection and dynamic modification give you a choice: you can build method wrappers, as this recipe does, in a smart and concise way; or, you can choose to code the boilerplate anyway, if you prefer to avoid what some would call the black magic of introspection and dynamic modification of class objects. The architecture of class list_with_aux_dict caters well to a rather common pattern of use, where sequence-modifying operations happen in bunches, followed by a period of time in which the sequence is not modified, but several membership tests may be performed. However, the addUnique_simple function shown earlier would not get any performance benefit if argument baseList was an instance of this recipe's list_with_aux_dict rather than a plain list: the function interleaves membership tests and sequence modifications. Therefore, too many rebuilds of the auxiliary dictionary for list_with_aux_dict would impede the function's performance. (Unless a typical case was for a vast majority of the items of otherList to be already contained in baseList, so that very few modifications occurred compared to the number of membership tests.) An important requisite for any of these membership-test optimizations is that the values in the sequence must be hashable (otherwise, of course, they cannot be keys in a dict, nor items in a set). For example, a list of tuples might be subjected to this recipe's treatment, but for a list of lists, the recipe as it stands is just not applicable. See Also The Library Reference and Python in a Nutshell sections on sequence types and mapping types. Recipe 5.13. Finding Subsequences Credit: David Eppstein, Alexander Semenov Problem You need to find occurrences of a subsequence in a larger sequence. Solution If the sequences are strings (plain or Unicode), Python strings' find method and the standard library's re module are the best approach. Otherwise, use the Knuth-Morris-Pratt algorithm (KMP): def KnuthMorrisPratt(text, pattern): ''' Yields all starting positions of copies of subsequence 'pattern' in sequence 'text' -- each argument can be any iterable. At the time of each yield, 'text' has been read exactly up to and including the match with 'pattern' that is causing the yield. ''' # ensure we can index into pattern, and also make a copy to protect # against changes to 'pattern' while we're suspended by `yield' pattern = list(pattern) length = len(pattern) # build the KMP "table of shift amounts" and name it 'shifts' shifts = [1] * (length + 1) shift = 1 for pos, pat in enumerate(pattern): while shift <= pos and pat != pattern[pos-shift]: shift += shifts[pos-shift] shifts[pos+1] = shift # perform the actual search startPos = 0 matchLen = 0 for c in text: while matchLen == length or matchLen >= 0 and pattern[matchLen] != c: startPos += shifts[matchLen] matchLen -= shifts[matchLen] matchLen += 1 if matchLen == length: yield startPos Discussion This recipe implements the Knuth-Morris-Pratt algorithm for finding copies of a given pattern as a contiguous subsequence of a larger text. Since KMP accesses the text sequentially, it is natural to implement it in a way that allows the text to be an arbitrary iterator. After a preprocessing stage that builds a table of shift amounts and takes time that's directly proportional to the length of the pattern, each text symbol is processed in constant amortized time. Explanations and demonstrations of how KMP works can be found in all good elementary texts about algorithms. (A recommendation is provided in See Also.)If text and pattern are both Python strings, you can get a faster solution by suitably applying Python built-in search methods: def finditer(text, pattern): pos = -1 while True: pos = text.find(pattern, pos+1) if pos < 0: break yield pos For example, using an alphabet of length 4 ('ACGU' . . .), finding all occurrences of a pattern of length 8 in a text of length 100000, on my machine, takes about 4.3 milliseconds with finditer, but the same task takes about 540 milliseconds with KnuthMorrisPratt (that's with Python 2.3; KMP is faster with Python 2.4, taking about 480 milliseconds, but that's still over 100 times slower than finditer). So remember: this recipe is useful for searches on generic sequences, including ones that you cannot keep in memory all at once, but if you're searching on strings, Python's built-in searching methods rule. See Also Many excellent books cover the fundamentals of algorithms; among such books, a widely admired one is Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, Clifford Stein, Introduction to Algorithms, 2d ed. (MIT Press). Recipe 5.14. Enriching the Dictionary Type with Ratings Functionality Credit: Dmitry Vasiliev, Alex Martelli Problem You want to use a dictionary to store the mapping between some keys and a current score value for each key. You frequently need to access the keys and scores in natural order (meaning, in order of ascending scores) and to check on a "key"'s current ranking in that order, so that using just a dict isn't quite enough. Solution We can subclass dict and add or override methods as needed. By using multiple inheritance, placing base UserDict.DictMixin before base dict and carefully arranging our various delegations and "over"rides, we can achieve a good balance between getting good performance and avoiding the need to write "boilerplate" code. By enriching our class with many examples in its docstring, we can use the standard library's module doctest to give us unit-testing functionality, as well as ensuring the accuracy of all the examples we write in the docstring: #!/usr/bin/env python ''' An enriched dictionary that holds a mapping from keys to scores ''' from bisect import bisect_left, insort_left import UserDict class Ratings(UserDict.DictMixin, dict): """ class Ratings is mostly like a dictionary, with extra features: the value corresponding to each key is the 'score' for that key, and all keys are ranked in terms their scores. Values must be comparable; keys, as well as being hashable, must be comparable if any two keys may ever have the same corresponding value (i.e., may be "tied" on score). All mapping-like behavior is just as you would expect, such as: >>> r = Ratings({"bob": 30, "john": 30}) >>> len(r) 2 >>> r.has_key("paul"), "paul" in r (False, False) >>> r["john"] = 20 >>> r.update({"paul": 20, "tom": 10}) >>> len(r) 4 >>> r.has_key("paul"), "paul" in r (True, True) >>> [r[key] for key in ["bob", "paul", "john", "tom"]] [30, 20, 20, 10] >>> r.get("nobody"), r.get("nobody", 0) (None, 0) In addition to the mapping interface, we offer rating-specific methods. r.rating(key) returns the ranking of a "key" in the ratings, with a ranking of 0 meaning the lowest score (when two keys have equal scores, the keys themselves are compared, to "break the tie", and the lesser key gets a lower ranking): >>> [r.rating(key) for key in ["bob", "paul", "john", "tom"]] [3, 2, 1, 0] getValueByRating(ranking) and getKeyByRating(ranking) return the score and key, respectively, for a given ranking index: >>> [r.getValueByRating(rating) for rating in range(4)] [10, 20, 20, 30] >>> [r.getKeyByRating(rating) for rating in range(4)] ['tom', 'john', 'paul', 'bob'] An important feature is that the keys( ) method returns keys in ascending order of ranking, and all other related methods return lists or iterators fully consistent with this ordering: >>> r.keys( ) ['tom', 'john', 'paul', 'bob'] >>> [key for key in r] ['tom', 'john', 'paul', 'bob'] >>> [key for key in r.iterkeys( )] ['tom', 'john', 'paul', 'bob'] >>> r.values( ) [10, 20, 20, 30] >>> [value for value in r.itervalues( )] [10, 20, 20, 30] >>> r.items( ) [('tom', 10), ('john', 20), ('paul', 20), ('bob', 30)] >>> [item for item in r.iteritems( )] [('tom', 10), ('john', 20), ('paul', 20), ('bob', 30)] An instance can be modified (adding, changing and deleting key-score correspondences), and every method of that instance reflects the instance's current state at all times: >>> r["tom"] = 100 >>> r.items( ) [('john', 20), ('paul', 20), ('bob', 30), ('tom', 100)] >>> del r["paul"] >>> r.items( ) [('john', 20), ('bob', 30), ('tom', 100)] >>> r["paul"] = 25 >>> r.items( ) [('john', 20), ('paul', 25), ('bob', 30), ('tom', 100)] >>> r.clear( ) >>> r.items( ) [ ] """ ''' the implementation carefully mixes inheritance and delegation to achieve reasonable performance while minimizing boilerplate, and, of course, to ensure semantic correctness as above. All mappings' methods not implemented below get inherited, mostly from DictMixin, but, crucially!, _ _getitem_ _ from dict. ''' def _ _init_ _(self, *args, **kwds): ''' This class gets instantiated just like 'dict' ''' dict._ _init_ _(self, *args, **kwds) # self._rating is the crucial auxiliary data structure: a list # of all (value, key) pairs, kept in "natural"ly-sorted order self._rating = [ (v, k) for k, v in dict.iteritems(self) ] self._rating.sort( ) def copy(self): ''' Provide an identical but independent copy ''' return Ratings(self) def _ _setitem_ _(self, k, v): ''' besides delegating to dict, we maintain self._rating ''' if k in self: del self._rating[self.rating(k)] dict._ _setitem_ _(self, k, v) insort_left(self._rating, (v, k)) def _ _delitem_ _(self, k): ''' besides delegating to dict, we maintain self._rating ''' del self._rating[self.rating(k)] dict._ _delitem_ _(self, k) ''' delegate some methods to dict explicitly to avoid getting DictMixin's slower (though correct) implementations instead ''' _ _len_ _ = dict._ _len_ _ _ _contains_ _ = dict._ _contains_ _ has_key = _ _contains_ _ ''' the key semantic connection between self._rating and the order of self.keys( ) -- DictMixin gives us all other methods 'for free', although we could implement them directly for slightly better performance. ''' def _ _iter_ _(self): for v, k in self._rating: yield k iterkeys = _ _iter_ _ def keys(self): return list(self) ''' the three ratings-related methods ''' def rating(self, key): item = self[key], key i = bisect_left(self._rating, item) if item == self._rating[i]: return i raise LookupError, "item not found in rating" def getValueByRating(self, rating): return self._rating[rating][0] def getKeyByRating(self, rating): return self._rating[rating][1] def _test( ): ''' we use doctest to test this module, which must be named rating.py, by validating all the examples in docstrings. ''' import doctest, rating doctest.testmod(rating) if _ _name_ _ == "_ _main_ _": _test( ) Discussion In many ways, a dictionary is the natural data structure for storing a correspondence between keys (e.g., names of contestants in a competition) and the current "score" of each key (e.g., the number of points a contestant has scored so far, or the highest bid made by each contestant at an auction, etc.). If we use a dictionary for such purposes, we will probably want to access it often in natural orderthe order in which the keys' scores are ascendingand we'll also want fast access to the rankings (ratings) implied by the current "score"s (e.g., the contestant currently in third place, the score of the contestant who is in second place, etc.). To achieve these purposes, this recipe subclasses dict to add the needed functionality that is completely missing from dict (methods rating, getValueByRating, getKeyByRating), and, more subtly and crucially, to modify method keys and all other related methods so that they return lists or iterators with the required order (i.e., the order in which scores are ascending; if we have to break ties when two keys have the same score, we implicitly compare the keys themselves). Most of the detailed documentation is in the docstring of the class itselfa crucial issue because bykeeping the documentation and examples there, we can use module doctest from the Python Standard Library to provide unit-testing functionality, as well as ensuring that our examples are correct. The most interesting aspect of the implementation is that it takes good care to minimize boilerplate (meaning repetitious and boring code, and therefore code where bugs are most likely to hide) without seriously impairing performance. class Ratings multiply inherits from dict and DictMixin, with the latter placed first in the list of bases, so that all methods come from the mixin, if it provides them, unless explicitly overridden in the class. Raymond Hettinger's DictMixin class was originally posted as a recipe to the online version of the Python Cookbook and later became part of Python 2.3's standard library. DictMixin provides all the methods of a mapping except _ _init_ _, copy, and the four fundamental methods: _ _getitem_ _, _ _setitem_ _, _ _delitem_ _, and, last but not least, keys. If you are coding a mapping class and want to ensure that your class supports all of the many methods that a full mapping provides to application code, you should subclass DictMixin and supply at least the fundamental methods (depending on your class' semanticse.g., if your class has immutable instances, you need not supply the mutator methods _ _setitem_ _ and _ _delitem_ _). You may optionally implement other methods for performance purposes, overriding the implementation that DictMixin provides. The whole DictMixin architecture can be seen as an excellent example of the classic Template Method Design Pattern, applied pervasively in a useful mix-in variant. In this recipe's class, we inherit _ _getitem_ _ from our other base (namely, the built-in type dict), and we also delegate explicitly to dict everything we can for performance reasons. We have to code the elementary mutator methods (_ _setitem_ _ and _ _delitem_ _) because, in addition to delegating to our base class dict, we need to maintain our auxiliary data structure self._ratinga list of (score, key) pairs that we keep in sorted order with the help of standard library module bisect. We implement keys ourselves (and while we're at it, we implement _ _iter_ _ i.e., iterkeys as well, since clearly keys is easiest to implement by using _ _iter_ _) to exploit self._rating and return the keys in the order we need. Finally, we add the obvious implementations for _ _init_ _ and copy, in addition to the three, ratings-specific methods that we supply. The result is quite an interesting example of balancing concision, clarity, and well-advised reuse of the enormous amount of functionality that the standard Python library places at our disposal. If you use this module in your applications, profiling may reveal that a method that this recipe's class inherits from DictMixin has somewhat unsatisfactory performanceafter all, the implementations in DictMixin are, of necessity, somewhat generic. If this is the case, by all means add a direct implementation of whatever further methods you need to achieve maximum performance! For example, if your application performs a lot of looping on the result of calling r.iteritems( ) for some instance r of class Ratings, you may get slightly better performance by adding to the body of the class the direct implementation of the method: def iteritems(self): for v, k in self._rating: yield k, v See Also Library Reference and Python in a Nutshell documentation about class DictMixin in module UserDict, and about module bisect. Recipe 5.15. Sorting Names and Separating Them by Initials Credit: Brett Cannon, Amos Newcombe Problem You want to write a directory for a group of people, and you want that directory to be grouped by the initials of their last names and sorted alphabetically. Solution Python 2.4's new itertools.groupby function makes this task easy: import itertools def groupnames(name_iterable): sorted_names = sorted(name_iterable, key=_sortkeyfunc) name_dict = { } for key, group in itertools.groupby(sorted_names, _groupkeyfunc): name_dict[key] = tuple(group) return name_dict pieces_order = { 2: (-1, 0), 3: (-1, 0, 1) } def _sortkeyfunc(name): ''' name is a string with first and last names, and an optional middle name or initial, separated by spaces; returns a string in order last-first-middle, as wanted for sorting purposes. ''' name_parts = name.split( ) return ' '.join([name_parts[n] for n in pieces_order[len(name_parts)]]) def _groupkeyfunc(name): ''' returns the key for grouping, i.e. the last name's initial. ''' return name.split( )[-1][0] Discussion In this recipe, name_iterable must be an iterable whose items are strings containing names in the form first - middle - last, with middle being optional and the parts separated by whitespace. The result of calling groupnames on such an iterable is a dictionary whose keys are the last names' initials, and the corresponding values are the tuples of all names with that last name's initial. Auxiliary function _sortkeyfunc splits a name that's a single string, either "first last" or "first middle last," and reorders the part into a list that starts with the last name, followed by first name, plus the middle name or initial, if any, at the end. Then, the function returns this list rejoined into a string. The resulting string is the key we want to use for sorting, according to the problem statement. Python 2.4's built-in function sorted takes just this kind of function (to call on each item to get the sort key) as the value of its optional parameter named key. Auxiliary function _groupkeyfunc takes a name in the same form and returns the last name's initialthe key on which, again according to the problem statement, we want to group.This recipe's primary function, groupnames, uses the two auxiliary functions and Python 2.4's sorted and itertools.groupby to solve our problem, building and returning the required dictionary. If you need to code this task in Python 2.3, you can use the same two support functions and recode function groupnames itself. In 2.3, it is more convenient to do the grouping first and the sorting separately on each group, since no groupby function is available in Python 2.3's standard library: def groupnames(name_iterable): name_dict = { } for name in name_iterable: key = _groupkeyfunc(name) name_dict.setdefault(key, [ ]).append(name) for k, v in name_dict.iteritems( ): aux = [(_sortkeyfunc(name), name) for name in v] aux.sort( ) name_dict[k] = tuple([ n for _ _, n in aux ]) return name_dict See Also Recipe 19.21; Library Reference (Python 2.4) docs on module itertools. Chapter 6. Object-Oriented Programming Introduction Recipe 6.1. Converting Among Temperature Scales Recipe 6.2. Defining Constants Recipe 6.3. Restricting Attribute Setting Recipe 6.4. Chaining Dictionary Lookups Recipe 6.5. Delegating Automatically as an Alternative to Inheritance Recipe 6.6. Delegating Special Methods in Proxies Recipe 6.7. Implementing Tuples with Named Items Recipe 6.8. Avoiding Boilerplate Accessors for Properties Recipe 6.9. Making a Fast Copy of an Object Recipe 6.10. Keeping References to Bound Methods Without Inhibiting Garbage Collection Recipe 6.11. Implementing a Ring Buffer Recipe 6.12. Checking an Instance for Any State Changes Recipe 6.13. Checking Whether an Object Has Necessary Attributes Recipe 6.14. Implementing the State Design Pattern Recipe 6.15. Implementing the "Singleton" Design Pattern Recipe 6.16. Avoiding the "Singleton" Design Pattern with the Borg Idiom Recipe 6.17. Implementing the Null Object Design Pattern Recipe 6.18. Automatically Initializing Instance Variables from _ _init_ _ Arguments Recipe 6.19. Calling a Superclass _ _init_ _ Method If It Exists Recipe 6.20. Using Cooperative Supercalls Concisely and Safely Introduction Credit: Alex Martelli, author of Python in a Nutshell (O'Reilly) Object-oriented programming (OOP) is among Python's greatest strengths. Python's OOP features continue to improve steadily and gradually, just like Python in general. You could already write better object-oriented programs in Python 1.5.2 (the ancient, long-stable version that was new when I first began to work with Python) than in any other popular language (excluding, of course, Lisp and its variants: I doubt there's anything you can't do well in Lisp-like languages, as long as you can stomach parentheses-heavy concrete syntax). For a few years now, since the release of Python 2.2, Python OOP has become substantially better than it was with 1.5.2. I am constantly amazed at the systematic progress Python achieves without sacrificing solidity, stability, and backwards-compatibility. To get the most out of Python's great OOP features, you should use them the Python way, rather than trying to mimic C++, Java, Smalltalk, or other languages you may be familiar with. You can do a lot of mimicry, thanks to Python's power. However, you'll get better mileage if you invest time and energy in understanding the Python way. Most of the investment is in increasing your understanding of OOP itself: what is OOP, what does it buy you, and which underlying mechanisms can your object-oriented programs use? The rest of the investment is in understanding the specific mechanisms that Python itself offers. One caveat is in order. For such a high-level language, Python is quite explicit about the OOP mechanisms it uses behind the curtains: they're exposed and available for your exploration and tinkering. Exploration and understanding are good, but beware the temptation to tinker. In other words, don't use unnecessary black magic just because you can. Specifically, don't use black magic in production code. If you can meet your goals with simplicity (and most often, in Python, you can), then keep your code simple. Simplicity pays off in readability, maintainability, and, more often than not, performance, too. To describe something as clever is not considered a compliment in the Python culture. So what is OOP all about? First of all, it's about keeping some state (data) and some behavior (code) together in handy packets. "Handy packets" is the key here. Every program has state and behaviorprogramming paradigms differ only in how you view, organize, and package them. If the packaging is in terms of objects that typically comprise state and behavior, you're using OOP. Some object-oriented languages force you to use OOP for everything, so you end up with many objects that lack either state or behavior. Python, however, supports multiple paradigms. While everything in Python is an object, you package things as OOP objects only when you want to. Other languages try to force your programming style into a predefined mold for your own good, while Python empowers you to make and express your own design choices. With OOP, once you have specified how an object is composed, you can instantiate as many objects of that kind as you need. When you don't want to create multiple objects, consider using other Python constructs, such as modules. In this chapter, you'll find recipes for Singleton, an object-oriented design pattern that eliminates the multiplicity of instantiation, and Borg, an idiom that makes multiple instances share state. But if you want only one instance, in Python it's often best to use a module, not an OOP object. To describe how an object is made, use the class statement: class SomeName(object): """ You usually define data and code here (in the class body). """ SomeName is a class object. It's a first-class object, like every Python object, meaning that you can put it in lists and dictionaries, pass it as an argument to functions, and so on. You don't haveto include the (object) part in the class header clauseclass SomeName: by itself is also valid Python syntaxbut normally you should include that part, as we'll see later. When you want a new instance of a class, call the class object as if it were a function. Each call returns a new instance object: anInstance = SomeName( ) another = SomeName( ) anInstance and another are two distinct instance objects, instances of the SomeName class. (See Recipe 4.18 for a class that does little more than this and yet is already quite useful.) You can freely bind (i.e., assign or set) and access (i.e., get) attributes (i.e., state) of an instance object: anInstance.someNumber = 23 * 45 print anInstance.someNumber # emits: 1035 Instances of an "empty" class like SomeName have no behavior, but they may have state. Most often, however, you want instances to have behavior. Specify the behavior you want by defining methods (with def statements, just like you define functions) inside the class body: class Behave(object): def _ _init_ _(self, name): self.name = name def once(self): print "Hello,", self.name def rename(self, newName) self.name = newName def repeat(self, N): for i in range(N): self.once( ) You define methods with the same def statement Python uses to define functions, exactly because methods are essentially functions. However, a method is an attribute of a class object, and its first formal argument is (by universal convention) named self. self always refers to the instance on which you call the method. The method with the special name _ _init_ _ is also known as the constructor (or more properly the initializer) for instances of the class. Python calls this special method to initialize each newly created instance with the arguments that you passed when calling the class (except for self, which you do not pass explicitly since Python supplies it automatically). The body of _ _init_ _ typically binds attributes on the newly created self instance to appropriately initialize the instance's state. Other methods implement the behavior of instances of the class. Typically, they do so by accessing instance attributes. Also, methods often rebind instance attributes, and they may call other methods. Within a class definition, these actions are always done with the self.something syntax. Once you instantiate the class, however, you call methods on the instance, access the instance's attributes, and even rebind them, using the theobject.something syntax: beehive = Behave("Queen Bee") beehive.repeat(3) beehive.rename("Stinger") beehive.once( ) print beehive.name beehive.name = 'See, you can rebind it "from the outside" too, if you want' beehive.repeat(2)self No true difference exists between what I described as the self.something syntax and the theobject.something syntax: the former is simply a special case of the latter, when the name of reference theobject happens to be self! If you're new to OOP in Python, you should try, in an interactive Python environment, the example snippets I have shown so far and those I'm going to show in the rest of this Introduction. One of the best interactive Python environments for such exploration is the GUI shell supplied as part of the free IDLE development environment that comes with Python. In addition to the constructor (_ _init_ _), your class may have other special methods, meaning methods with names that start and end with two underscores. Python calls the special methods of a class when instances of the class are used in various operations and built-in functions. For example, len(x) returns x._ _len_ _( ); a+b normally returns a._ _add_ _(b); a[b] returns a._ _getitem_ _(b). Therefore, by defining special methods in a class, you can make instances of that class interchangeable with objects of built-in types, such as numbers, lists, and dictionaries. Each operation and built-in function can try several special methods in some specific order. For example, a+b first tries a._ _add_ _(b), but, if that doesn't pan out, the operation also gives object b a say in the matter, by next trying b._ _radd_ _(a). This kind of intrinsic structuring among special methods, that operations and built-in functions can provide, is an important added value of such functions and operations with respect to pure OO notation such as someobject.somemethod(arguments). The ability to handle different objects in similar ways, known as polymorphism, is a major advantage of OOP. Thanks to polymorphism, you can call the same method on various objects, and each object can implement the method appropriately. For example, in addition to the Behave class, you might have another class that implements a repeat method with rather different behavior: class Repeater(object): def repeat(self, N): print N*"*-*" You can mix instances of Behave and Repeater at will, as long as the only method you call on each such instance is repeat: aMix = beehive, Behave('John'), Repeater( ), Behave('world') for whatever in aMix: whatever.repeat(3) Other languages require inheritance, or the formal definition and implementation of interfaces, in order to enable such polymorphism. In Python, all you need is to have methods with the same signature (i.e., methods of the same name, callable with the same arguments). This signature- based polymorphism allows a style of programming that's quite similar to generic programming (e.g., as supported by C++'s template classes and functions), without syntax cruft and without conceptual complications. Python also uses inheritance, which is mostly a handy, elegant, structured way to reuse code. You can define a class by inheriting from another (i.e., subclassing the other class) and thenadding or redefining (known as overriding) some methods: class Subclass(Behave): def once(self): print '(%s)' % self.name subInstance = Subclass("Queen Bee") subInstance.repeat(3) The Subclass class overrides only the once method, but you can also call the repeat method on subInstance, since Subclass inherits that method from the Behave superclass. The body of the repeat method calls once n times on the specific instance, using whatever version of the once method the instance has. In this case, each call uses the method from the Subclass class, which prints the name in parentheses, not the original version from the Behave class, which prints the name after a greeting. The idea of a method calling other methods on the same instance and getting the appropriately overridden version of each is important in every object-oriented language, including Python. It is also known as the Template Method Design Pattern. The method of a subclass often overrides a method from the superclass, but also needs to call the method of the superclass as part of its own operation. You can do this in Python by explicitly getting the method as a class attribute and passing the instance as the first argument: class OneMore(Behave): def repeat(self, N): Behave.repeat(self, N+1) zealant = OneMore("Worker Bee") zealant.repeat(3) The OneMore class implements its own repeat method in terms of the method with the same name in its superclass, Behave, with a slight change. This approach, known as delegation, is pervasive in all programming. Delegation involves implementing some functionality by letting another existing piece of code do most of the work, often with some slight variation. An overriding method often is best implemented by delegating some of the work to the same method in the superclass. In Python, the syntax Classname.method(self, . . .) delegates to Classname's version of the method. A vastly preferable way to perform superclass delegation, however, is to use Python's built-in super: class OneMore(Behave): def repeat(self, N): super(OneMore, self).repeat(N+1) This super construct is equivalent to the explicit use of Behave.repeat in this simple case, but it also allows class OneMore to be used smoothly with multiple inheritance. Even if you're not interested in multiple inheritance at first, you should still get into the habit of using super instead of explicit delegation to your base class by namesuper costs nothing and it may prove very useful to you in the future. Python does fully support multiple inheritance: one class can inherit from several other classes. In terms of coding, this feature is sometimes just a minor one that lets you use the mix-in class idiom, a convenient way to supply functionality across a broad range of classes. (See Recipe 6.20 and Recipe 6.12, for unusual but powerful examples of using the mix-in idiom.) However, multiple inheritance is particularly important because of its implications for object-oriented analysisthe way you conceptualize your problem and your solution in the first place. Single inheritance pushes you to frame your problem space via taxonomy (i.e., mutually exclusive classification). The real world doesn't work like that. Rather, it resembles Jorge Luis Borges' explanation in The Analytical Language of John Wilkins, from a purported Chinese encyclopedia, The Celestial Emporium of Benevolent Knowledge. Borges explains that all animals are divided into: Those that belong to the EmperorEmbalmed ones Those that are trained Suckling pigs Mermaids Fabulous ones Stray dogs Those included in the present classification Those that tremble as if they were mad Innumerable ones Those drawn with a very fine camelhair brush Others Those that have just broken a flower vase Those that from a long way off look like flies You get the point: taxonomy forces you to pigeonhole, fitting everything into categories that aren't truly mutually exclusive. Modeling aspects of the real world in your programs is hard enough without buying into artificial constraints such as taxonomy. Multiple inheritance frees you from these constraints. Ah, yes, that (object) thingI had promised to come back to it later. Now that you've seen Python's notation for inheritance, you realize that writing class X(object) means that class X inherits from class object. If you just write class Y:, you're saying that Y doesn't inherit from anythingY, so to speak, "stands on its own". For backwards compatibility, Python allows you to request such a rootless class, and, if you do, then Python makes class Y an "old-style" class, also known as a classic class, meaning a class that works just like all classes used to work in the Python versions of old. Python is very keen on backwards-compatibility. For many elementary uses, you won't notice the difference between classic classes and the new- style classes that are recommended for all new Python code you write. However, it's important to underscore that classic classes are a legacy feature, not recommended for new code. Even within the limited compass of elementary OOP features that I cover in this Introduction, you will already feel some of the limitations of classic classes: for example, you cannot use super within classic classes, and in practice, you should not do any serious use of multiple inheritance with them. Many important features of today's Python OOP, such as the property built-in, can't work completely, if they even work at all, with old-style classes. In practice, even if you're maintaining a large body of legacy Python code, the next time you need to do any substantial maintenance on that code, you should take the little effort required to ensure all classes are new style: it's a small job, and it will ease your future maintenance burden quite a bit. Instead of explicitly having all your classes inherit from object, an equivalent alternative is to add the following assignment statement close to the start of every module that defines any classes: _ _metaclass_ _ = type The built-in type is the metaclass of object and of every other new-style class and built-in type. That's why inheriting from object or any built-in type makes a class new style: the class you're coding gets the same metaclass as its base. A class without bases can get its metaclass from the module-global _ _metaclass_ _ variable, which is why the "state"ment I suggest suffices toensure that any classes without explicit bases are made new-style. Even if you never make any other use of explicit metaclasses (a rather advanced subject that is, nevertheless, mentioned in several of this chapter's recipes), this one simple use of them will stand you in good stead. What Is a Metaclass? Metaclasses do not mean "deep, dark black magic". When you execute any class statement, Python performs the following steps: Remember the class name as a string, say n, and the class bases as a tuple, say b. Execute the body of the class, recording all names that the body binds as keys in a new dictionary d, each with its associated value (e.g., each statement such as def f(self) just sets d['f'] to the function object the def statement builds). Determine the appropriate metaclass, say M, by inheritance or by looking for name _ _metaclass_ _ in d and in the globals: if '_ _metaclass_ _' in d: M = d['_ _metaclass_ _'] elif b: M = type(b[0]) elif '_ _metaclass_ _' in globals( ): M = globals( )['_ _metaclass_ _'] else: M = types.ClassType types.ClassType is the metaclass of old-style classes, so this code implies that a class without bases is old style if the name '_ _metaclass_ _' is not set in the class body nor among the global variables of the current module. Call M(n, b, d) and record the result as a variable with name n in whatever scope the class statement executed. So, some metaclass M is always involved in the execution of any class statement. The metaclass is normally type for new-style classes, types.ClassType for old-style classes. You can set it up to use your own custom metaclass (normally a subclass of type), and that is where you may reasonably feel that things are getting a bit too advanced. However, understanding that a class statement, such as: class Someclass(Somebase): _ _metaclass_ _ = type x = 23 is exactly equivalent to the assignment statement: Someclass = type('Someclass', (Somebase,), {'x': 23}) does help a lot in understanding the exact semantics of the class statement. Recipe 6.1. Converting Among Temperature Scales Credit: Artur de Sousa Rocha, Adde Nilsson Problem You want to convert easily among Kelvin, Celsius, Fahrenheit, and Rankine scales of temperature. Solution Rather than having a dozen functions to do all possible conversions, we can more elegantly package this functionality into a class: class Temperature(object): coefficients = {'c': (1.0, 0.0, -273.15), 'f': (1.8, -273.15, 32.0), 'r': (1.8, 0.0, 0.0)} def _ _init_ _(self, **kwargs): # default to absolute (Kelvin) 0, but allow one named argument, # with name being k, c, f or r, to use any of the scales try: name, value = kwargs.popitem( ) except KeyError: # no arguments, so default to k=0 name, value = 'k', 0 # error if there are more arguments, or the arg's name is unknown if kwargs or name not in 'kcfr': kwargs[name] = value # put it back for diagnosis raise TypeError, 'invalid arguments %r' % kwargs setattr(self, name, float(value)) def _ _getattr_ _(self, name): # maps getting of c, f, r, to computation from k try: eq = self.coefficients[name] except KeyError: # unknown name, give error message raise AttributeError, name return (self.k + eq[1]) * eq[0] + eq[2] def _ _setattr_ _(self, name, value): # maps settings of k, c, f, r, to setting of k; forbids others if name in self.coefficients: # name is c, f or r -- compute and set k eq = self.coefficients[name] self.k = (value - eq[2]) / eq[0] - eq[1] elif name == 'k': # name is k, just set it object._ _setattr_ _(self, name, value) else: # unknown name, give error message raise AttributeError, name def _ _str_ _(self): # readable, concise representation as string return "%s K" % self.k def _ _repr_ _(self): # detailed, precise representation as string return "Temperature(k=%r)" % self.k Discussion Converting between several different scales or units of measure is a task that's subject to a "combinatorial explosion": if we tackle it in the apparently obvious way, by providing a function for each conversion, then, to deal with n different units, we will have to write n * (n-1) functions. A Python class can intercept attribute setting and getting, and perform computation on the fly in response. This power enables a much handier and more elegant architecture, as shown in this recipe for the specific case of temperatures. Inside the class, we always hold the measurement in one reference unit or scale, Kelvin (absolute) degrees in the case of this recipe. We allow the setting of the value to happen through any of four attribute names ('k', 'r', 'c', 'f', abbreviations of the scales' names), and compute and set the Kelvin-scale value appropriately. Vice versa, we also allow the "getting" of the value in any scale, through the same attribute names, computing the result on the fly. (Assuming you have saved the code in this recipe as te.py somewhere on your Python sys.path, you can import it as a module.) For example: >>> from te import Temperature >>> t = Temperature(f=70) # 70 F is... >>> print t.c # ...a bit over 21 C 21.1111111111 >>> t.c = 23 # 23 C is... >>> print t.f # ...a bit over 73 F 73.4 _ _getattr_ _ and _ _setattr_ _ work better than named properties would in this case, since the form of the computation is the same for every attribute (except the reference 'k' one), and we only need to use different coefficients that we can most handily keep in a per-class dictionary, the one we name self.coefficients. It's important to remember that _ _setattr_ _ is called on every setting of any attribute, so it must delegate to object the setting of attributes, which need to be recorded in the instance (the _ _setattr_ _ implementation in this recipe does just such a delegation for attribute k) and must raise an AttributeError exception for attributes that can't be set. _ _getattr_ _, on the other hand, is called only upon the "getting" of an attribute that can't be found by other, "normal" means (e.g., in the case of this recipe's class, _ _getattr_ _ is not called for accesses to attribute k, which is recorded in the instance and thus gets found by normal means). _ _getattr_ _ must also raise an AttributeError exception for attributes that can't be accessed. See Also Library Reference and Python in a Nutshell documentation on attributes and on special methods _ _getattr_ _ and _ _setattr_ _. Recipe 6.2. Defining Constants Credit: Alex Martelli Problem You need to define module-level variables (i.e., named constants) that client code cannot accidentally rebind. Solution You can install any object as if it were a module. Save the following code as module const.py on some directory on your Python sys.path: class _const(object): class ConstError(TypeError): pass def _ _setattr_ _(self, name, value): if name in self._ _dict_ _: raise self.ConstError, "Can't rebind const(%s)" % name self._ _dict_ _[name] = value def _ _delattr_ _(self, name): if name in self._ _dict_ _: raise self.ConstError, "Can't unbind const(%s)" % name raise NameError, name import sys sys.modules[_ _name_ _] = _const( ) Now, any client code can import const, then bind an attribute on the const module just once, as follows: const.magic = 23 Once the attribute is bound, the program cannot accidentally rebind or unbind it: const.magic = 88 # raises const.ConstError del const.magic # raises const.ConstError Discussion In Python, variables can be rebound at will, and modules, differently from classes, don't let you define special methods such as _ _setattr_ _ to stop rebinding. An easy solution is to install an instance as if it were a module. Python performs no type-checks to force entries in sys.modules to actually be module objects. Therefore, you can install any object there and take advantage of attribute-access special methods (e.g., to prevent rebinding, to synthesize attributes on the fly in _ _getattr_ _, etc.), while still allowing client code to access the object with import somename. You may even see it as a more Pythonic Singleton-style idiom (but see Recipe 6.16).This recipe ensures that a module-level name remains constantly bound to the same object once it has first been bound to it. This recipe does not deal with a certain object's immutability, which is quite a different issue. Altering an object and rebinding a name are different concepts, as explained in Recipe 4.1. Numbers, strings, and tuples are immutable: if you bind a name in const to such an object, not only will the name always be bound to that object, but the object's contents also will always be the same since the object is immutable. However, other objects, such as lists and dictionaries, are mutable: if you bind a name in const to, say, a list object, the name will always remain bound to that list object, but the contents of the list may change (e.g., items in it may be rebound or unbound, more items can be added with the object's append method, etc.). To make "read-only" wrappers around mutable objects, see Recipe 6.5. You might choose to have class _const's _ _setattr_ _ method perform such wrapping implicitly. Say you have saved the code from Recipe 6.5 as module ro.py somewhere along your Python sys.path. Then, you need to add, at the start of module const.py: import ro and change the assignment self._ _dict_ _[name] = value, used in class _const's _ _setattr_ _ method to: self._ _dict_ _[name] = ro.Readonly(value) Now, when you set an attribute in const to some value, what gets bound there is a read-only wrapper to that value. The underlying value might still get changed by calling mutators on some other reference to that same value (object), but it cannot be accidentally changed through the attribute of "pseudo-module" const. If you want to avoid such "accidental changes through other references", you need to take a copy, as explained in Recipe 4.1, so that there exist no other references to the value held by the read-only wrapper. Ensure that at the start of module const.py you have: import ro, copy and change the assignment in class _const's _ _setattr_ _ method to: self._ _dict_ _[name] = ro.Readonly(copy.copy(value)) If you're sufficiently paranoid, you might even use copy.deepcopy rather than plain copy.copy in this latest snippet. However, you may end up paying substantial amounts of memory, as well as losing some performance, by these kinds of excessive precautions. You should evaluate carefully whether so much prudence is really necessary for your specific application. Whatever you end up deciding about this issue, Python offers all the tools you need to implement exactly the amount of constantness you require. The _const class presented in this recipe can be seen, in a sense, as the "complement" of the NoNewAttrs class, which is presented next in Recipe 6.3. This one ensures that already bound attributes can never be rebound but lets you freely bind new attributes; the other one, conversely, lets you freely rebind attributes that are already bound but blocks the binding of any new attribute. See Also Recipe 6.5; Recipe 6.13; Recipe 4.1; Library Reference and Python in a Nutshell docs on module objects, the import statement, and the modules attribute of the sys built-in module. Recipe 6.3. Restricting Attribute Setting Credit: Michele Simionato Problem Python normally lets you freely add attributes to classes and their instances. However, you want to restrict that freedom for some class. Solution Special method _ _setattr_ _ intercepts every setting of an attribute, so it lets you inhibit the addition of new attributes that were not already present. One elegant way to implement this idea is to code a class, a simple custom metaclass, and a wrapper function, all cooperating for the purpose, as follows: def no_new_attributes(wrapped_setattr): """ raise an error on attempts to add a new attribute, while allowing existing attributes to be set to new values. """ def _ _setattr_ _(self, name, value): if hasattr(self, name): # not a new attribute, allow setting wrapped_setattr(self, name, value) else: # a new attribute, forbid adding it raise AttributeError("can't add attribute %r to %s" % (name, self)) return _ _setattr_ _ class NoNewAttrs(object): """ subclasses of NoNewAttrs inhibit addition of new attributes, while allowing existing attributed to be set to new values. """ # block the addition new attributes to instances of this class _ _setattr_ _ = no_new_attributes(object._ _setattr_ _) class _ _metaclass_ _(type): " simple custom metaclass to block adding new attributes to this class " _ _setattr_ _ = no_new_attributes(type._ _setattr_ _) Discussion For various reasons, you sometimes want to restrict Python's dynamism. In particular, you may want to get an exception when a new attribute is accidentally set on a certain class or one of its instances. This recipe shows how to go about implementing such a restriction. The key point of the recipe is, don't use _ _slots_ _ for this purpose: _ _slots_ _ is intended for a completely different task (i.e., saving memory by avoiding each instance having a dictionary, as it normally would, when you need to have vast numbers of instances of a class with just a few fixed attributes). _ _slots_ _ performs its intended task well but has various limitations when you try to stretch it to perform, instead, the task this recipe covers. (See Recipe 6.18 for an example of the appropriate use of _ _slots_ _ to save memory.) Notice that this recipe inhibits the addition of runtime attributes, not only to class instances, but also to the class itself, thanks to the simple custom metaclass it defines. When you want toinhibit accidental addition of attributes, you usually want to inhibit it on the class as well as on each individual instance. On the other hand, existing attributes on both the class and its instances may be freely set to new values. Here is an example of how you could use this recipe: class Person(NoNewAttrs): firstname = '' lastname = '' def _ _init_ _(self, firstname, lastname): self.firstname = firstname self.lastname = lastname def _ _repr_ _(self): return 'Person(%r, %r)' % (self.firstname, self.lastname) me = Person("Michere", "Simionato") print me # emits: Person('Michere', 'Simionato') # oops, wrong value for firstname, can we fix it? Sure, no problem! me.firstname = "Michele" print me # emits: Person('Michele', 'Simionato') The point of inheriting from NoNewAttrs is forcing yourself to "declare" all allowed attributes by setting them at class level in the body of the class itself. Any further attempt to set a new, "undeclared" attribute raises an AttributeError: try: Person.address = '' except AttributeError, err: print 'raised %r as expected' % err try: me.address = '' except AttributeError, err: print 'raised %r as expected' % err In some ways, therefore, subclasses of NoNewAttr and their instances behave more like Java or C++ classes and instances, rather than normal Python ones. Thus, one use case for this recipe is when you're coding in Python a prototype that you already know will eventually have to be recoded in a less dynamic language. See Also Library Reference and Python in a Nutshell documentation on the special method _ _setattr_ _ and on custom metaclasses; Recipe 6.18 for an example of an appropriate use of _ _slots_ _ to save memory; Recipe 6.2 for a class that is the complement of this one. Recipe 6.4. Chaining Dictionary Lookups Credit: Raymond Hettinger Problem You have several mappings (usually dicts) and want to look things up in them in a chained way (try the first one; if the key is not there, then try the second one; and so on). Specifically, you want to make a single mapping object that "virtually merges" several others, by looking things up in them in a specified priority order, so that you can conveniently pass that one object around. Solution A mapping is a generalized, abstract version of a dictionary: a mapping provides an interface that's similar to a dictionary's, but it may use very different implementations. All dictionaries are mappings, but not vice versa. Here, you need to implement a mapping which sequentially tries delegating lookups to other mappings. A class is the right way to encapsulate this functionality: class Chainmap(object): def _ _init_ _(self, *mappings): # record the sequence of mappings into which we must look self._mappings = mappings def _ _getitem_ _(self, key): # try looking up into each mapping in sequence for mapping in self._mappings: try: return mapping[key] except KeyError: pass # `key' not found in any mapping, so raise KeyError exception raise KeyError, key def get(self, key, default=None): # return self[key] if present, otherwise `default' try: return self[key] except KeyError: return default def _ _contains_ _(self, key): # return True if `key' is present in self, otherwise False try: self[key] return True except KeyError: return False For example, you can now implement the same sequence of lookups that Python normally uses for any name: look among locals, then (if not found there) among globals, lastly (if not found yet) among built-ins: import _ _builtin_ _pylookup = Chainmap(locals( ), globals( ), vars(_ _builtin_ _)) Discussion Chainmap relies on minimal functionality from the mappings it wraps: each of those underlying mappings must allow indexing (i.e., supply a special method _ _getitem_ _), and it must raise the standard exception KeyError when indexed with a key that the mapping does not know about. A Chainmap instance provides the same behavior, plus the handy get method covered in Recipe 4.9 and special method _ _contains_ _ (which conveniently lets you check whether some key k is present in a Chainmap instance c by just coding if k in c). Besides the obvious and sensible limitation of being "read-only", this Chainmap class has othersessentially, it is not a "full mapping" even within the read-only design choice. You can make any partial mapping into a "full mapping" by inheriting from class DictMixin (in standard library module UserDict) and supplying a few key methods (DictMixin implements the others). Here is how you could make a full (read-only) mapping from ChainMap and UserDict.DictMixin: import UserDict from sets import Set class FullChainmap(Chainmap, UserDict.DictMixin): def copy(self): return self._ _class_ _(self._mappings) def _ _iter_ _(self): seen = Set( ) for mapping in self._mappings: for key in mapping: if key not in seen: yield key seen.add(key) iterkeys = _ _iter_ _ def keys(self): return list(self) This class FullChainmap adds one requirement to the mappings it holds, besides the requirements posed by Chainmap: the mappings must be iterable. Also note that the implementation in Chainmap of methods get and _ _contains_ _ is redundant (although innocuous) once we subclass DictMixin, since DictMixin also implements those two methods (as well as many others) in terms of lower-level methods, just like Chainmap does. See Recipe 5.14 for more details about DictMixin. See Also Recipe 4.9; Recipe 5.14; the Library Reference and Python in a Nutshell sections on mapping types. Recipe 6.5. Delegating Automatically as an Alternative to Inheritance Credit: Alex Martelli, Raymond Hettinger Problem You'd like to inherit from a class or type, but you need some tweak that inheritance does not provide. For example, you want to selectively hide some of the base class' methods, which inheritance doesn't allow. Solution Inheritance is quite handy, but it's not all-powerful. For example, it doesn't let you hide methods or other attributes supplied by a base class. Containment with automatic delegation is often a good alternative. Say, for example, you need to wrap some objects to make them read-only; thus preventing accidental alterations. Therefore, besides stopping attribute-setting, you also need to hide mutating methods. Here's a way: # support 2.3 as well as 2.4 try: set except NameError: from sets import Set as set class ROError(AttributeError): pass class Readonly: # there IS a reason to NOT subclass object, see Discussion mutators = { list: set('''_ _delitem_ _ _ _delslice_ _ _ _iadd_ _ _ _imul_ _ _ _setitem_ _ _ _setslice_ _ append extend insert pop remove sort'''.split( )), dict: set('''_ _delitem_ _ _ _setitem_ _ clear pop popitem setdefault update'''.split( )), } def _ _init_ _(self, o): object._ _setattr_ _(self, '_o', o) object._ _setattr_ _(self, '_no', self.mutators.get(type(o), ( ))) def _ _setattr_ _(self, n, v): raise ROError, "Can't set attr %r on RO object" % n def _ _delattr_ _(self, n): raise ROError, "Can't del attr %r from RO object" % n def _ _getattr_ _(self, n): if n in self._no: raise ROError, "Can't get attr %r from RO object" % n return getattr(self._o, n) Code using this class Readonly can easily add other wrappable types with Readonly.mutators[sometype] = the_mutators. Discussion Automatic delegation, which the special methods _ _getattr_ _, _ _setattr_ _, and __delattr_ _ enable us to perform so smoothly, is a powerful, general technique. In this recipe, we show how to use it to get an effect that is almost indistinguishable from subclassing while hiding some names. In particular, we apply this quasi-subclassing to the task of wrapping objects to make them read-only. Performance isn't quite as good as it might be with real inheritance, but we get better flexibility and finer-grained control as compensation. The fundamental idea is that each instance of our class holds an instance of the type we are wrapping (i.e., extending and/or tweaking). Whenever client code tries to get an attribute from an instance of our class, unless the attribute is specifically defined there (e.g., the mutators dictionary in class Readonly), _ _getattr_ _ TRansparently shunts the request to the wrapped instance after appropriate checks. In Python, methods are also attributes, accessed in just the same way, so we don't need to do anything different to access methods. The _ _getattr_ _ approach used to access data attributes works for methods just as well. This is where the comment in the recipe about there being a specific reason to avoid subclassing object comes in. Our _ _getattr_ _ based approach does work on special methods too, but only for instances of old-style classes. In today's object model, Python operations access special methods on the class, not on the instance. Solutions to this issue are presented next in Recipe 6.6 and in Recipe 20.8. The approach adopted in this recipemaking class Readonly old style, so that the issue can be locally avoided and delegated to other recipesis definitely not recommended for production code. I use it here only to keep this recipe shorter and to avoid duplicating coverage that is already amply given elsewhere in this cookbook. _ _setattr_ _ plays a role similar to _ _getattr_ _, but it gets called when client code sets an instance attribute; in this case, since we want to make a read-only wrapper, we simply forbid the operation. Remember, to avoid triggering _ _setattr_ _ from inside the methods you code, you must never code normal self.n = v statements within the methods of classes that have _ _setattr_ _. The simplest workaround is to delegate the setting to class object, just like our class Readonly does twice in its _ _init_ _ method. Method _ _delattr_ _ completes the picture, dealing with any attempts to delete attributes from an instance. Wrapping by automatic delegation does not work well with client or framework code that, one way or another, does type-testing. In such cases, the client or framework code is breaking polymorphism and should be rewritten. Remember not to use type-tests in your own client code, as you probably do not need them anyway. See Recipe 6.13 for better alternatives. In old versions of Python, automatic delegation was even more prevalent, since you could not subclass built-in types. In modern Python, you can inherit from built-in types, so you'll use automatic delegation less often. However, delegation still has its placeit is just a bit farther from the spotlight. Delegation is more flexible than inheritance, and sometimes such flexibility is invaluable. In addition to the ability to delegate selectively (thus effectively "hiding" some of the attributes), an object can delegate to different subobjects over time, or to multiple subobjects at one time, and inheritance doesn't offer anything comparable. Here is an example of delegating to multiple specific subobjects. Say that you have classes that are chock full of "forwarding methods", such as: class Pricing(object): def _ _init_ _(self, location, event): self.location = location self.event = event def setlocation(self, location): self.location = location def getprice(self): return self.location.getprice( ) def getquantity(self): return self.location.getquantity( ) def getdiscount(self): return self.event.getdiscount( ) and many more such methodsInheritance is clearly not applicable because an instance of Pricing must delegate to specific location and event instances, which get passed at initialization time and may even be changed. Automatic delegation to the rescue: class AutoDelegator(object): delegates = ( ) do_not_delegate = ( ) def _ _getattr_ _(self, key): if key not in do_not_delegate: for d in self.delegates: try: return getattr(d, key) except AttributeError: pass raise AttributeError, key class Pricing(AutoDelegator): def _ _init_ _(self, location, event): self.delegates = [location, event] def setlocation(self, location): self.delegates[0] = location In this case, we do not delegate the setting and deletion of attributes, only the getting of attributes (and nonspecial methods). Of course, this approach is fully applicable only when the methods (and other attributes) of the various objects to which we want to delegate do not interfere with each other; for example, location must not have a getdiscount method; otherwise, it would preempt the delegation of that method, which is intended to go to event. If a class that does lots of delegation has a few such issues to solve, it can do so by explicitly defining the few corresponding methods, since _ _getattr_ _ enters the picture only for attributes and methods that cannot be found otherwise. The ability to hide some attributes and methods that are supplied by a delegate, but the delegator does not want to expose, is supported through attribute do_not_delegate, which any subclass may override. For example, if class Pricing wanted to hide a method setdiscount that is supplied by, say, event, only a tiny change would be required: class Pricing(AutoDelegator): do_not_delegate = ('set_discount',) while all the rest remains as in the previous snippet. See Also Recipe 6.13; Recipe 6.6; Recipe 20.8; Python in a Nutshell chapter on OOP; PEP 253 (http://www.python.org/peps/pep-0253.html) for more details about Python's current (new- style) object model. Recipe 6.6. Delegating Special Methods in Proxies Credit: Gonçalo Rodrigues Problem In the new-style object model, Python operations perform implicit lookups for special methods on the class (rather than on the instance, as they do in the classic object model). Nevertheless, you need to wrap new-style instances in proxies that can also delegate a selected set of special methods to the object they're wrapping. Solution You need to generate each proxy's class on the fly. For example: class Proxy(object): """ base class for all proxies """ def _ _init_ _(self, obj): super(Proxy, self)._ _init_ _(obj) self._obj = obj def _ _getattr_ _(self, attrib): return getattr(self._obj, attrib) def make_binder(unbound_method): def f(self, *a, **k): return unbound_method(self._obj, *a, **k) # in 2.4, only: f._ _name_ _ = unbound_method._ _name_ _ return f known_proxy_classes = { } def proxy(obj, *specials): ''' factory-function for a proxy able to delegate special methods ''' # do we already have a suitable customized class around? obj_cls = obj._ _class_ _ key = obj_cls, specials cls = known_proxy_classes.get(key) if cls is None: # we don't have a suitable class around, so let's make it cls = type("%sProxy" % obj_cls._ _name_ _, (Proxy,), { }) for name in specials: name = '_ _%s_ _' % name unbound_method = getattr(obj_cls, name) setattr(cls, name, make_binder(unbound_method)) # also cache it for the future known_proxy_classes[key] = cls # instantiate and return the needed proxy return cls(obj) Discussion Proxying and automatic delegation are a joy in Python, thanks to the _ _getattr_ _ hook. Python calls it automatically when a lookup for any attribute (including a methodPython draws no distinction there) has not otherwise succeeded.In the old-style (classic) object model, _ _getattr_ _ also applied to special methods that were looked up as part of a Python operation. This required some care to avoid mistakenly supplying a special method one didn't really want to supply but was otherwise handy. Nowadays, the new- style object model is recommended for all new code: it is faster, more regular, and richer in features. You get new-style classes when you subclass object or any other built-in type. One day, some years from now, Python 3.0 will eliminate the classic object model, as well as other features that are still around only for backwards-compatibility. (See http://www.python.org/peps/pep-3000.html for details about plans for Python 3.0almost all changes will be language simplifications, rather than new features.) In the new-style object model, Python operations don't look up special methods at runtime: they rely on "slots" held in class objects. Such slots are updated when a class object is built or modified. Therefore, a proxy object that wants to delegate some special methods to an object it's wrapping needs to belong to a specially made and tailored class. Fortunately, as this recipe shows, making and instantiating classes on the fly is quite an easy job in Python. In this recipe, we don't use any advanced Python concepts such as custom metaclasses and custom descriptors. Rather, each proxy is built by a factory function proxy, which takes as arguments the object to wrap and the names of special methods to delegate (shorn of leading and trailing double underscores). If you've saved the "Solution"'s code in a file named proxy.py somewhere along your Python sys.path, here is how you could use it from an interactive Python interpreter session: >>> import proxy >>> a = proxy.proxy([ ], 'len', 'iter') # only delegate _ _len_ _ & _ _iter_ _ >>> a # _ _repr_ _ is not delegated >>> a._ _class_ _ >>> a._obj [ ] >>> a.append # all non-specials are delegated Since _ _len_ _ is delegated, len(a) works as expected: >>> len(a) 0 >>> a.append(23) >>> len(a) 1 Since _ _iter_ _ is delegated, for loops work as expected, as does intrinsic looping performed by built-ins such as list, sum, max, . . . : >>> for x in a: print x ... 23 >>> list(a) [23] >>> sum(a) 23 >>> max(a) 23 However, since _ _getitem_ _ is not delegated, a cannot be indexed nor sliced: >>> a._ _getitem_ _ >>> a[1] Traceback (most recent call last): File "", line 1, in ? TypeError: unindexable object Function proxy uses a "cache" of classes it has previously generated, the global dictionary known_proxy_classes, keyed by the class of the object being wrapped and the tuple of special methods' names being delegated. To make a new class, proxy calls the built-in type, passing as arguments the name of the new class (made by appending 'Proxy' to the name of the class being wrapped), class Proxy as the only base, and an "empty" class dictionary (since it's adding no class attributes yet). Base class Proxy deals with initialization and delegation of ordinary attribute lookups. Then, factory function proxy loops over the names of specials to be delegated: for each of them, it gets the unbound method from the class of the object being wrapped, and sets it as an attribute of the new class within a make_binder closure. make_binder deals with calling the unbound method with the appropriate first argument (i.e., the object being wrapped, self._obj). Once it's done preparing a new class, proxy saves it in known_proxy_classes under the appropriate key. Finally, whether the class was just built or recovered from known_proxy_classes, proxy instantiates it, with the object being wrapped as the only argument, and returns the resulting proxy instance. See Also Recipe 6.5 for more information about automatic delegation; Recipe 6.9 for another example of generating classes on the fly (using a class statement rather than a call to type). Recipe 6.7. Implementing Tuples with Named Items Credit: Gonçalo Rodrigues, Raymond Hettinger Problem Python tuples are handy ways to group pieces of information, but having to access each item by numeric index is a bother. You'd like to build tuples whose items are also accessible as named attributes. Solution A factory function is the simplest way to generate the required subclass of tuple: # use operator.itemgetter if we're in 2.4, roll our own if we're in 2.3 try: from operator import itemgetter except ImportError: def itemgetter(i): def getter(self): return self[i] return getter def superTuple(typename, *attribute_names): " create and return a subclass of `tuple', with named attributes " # make the subclass with appropriate _ _new_ _ and _ _repr_ _ specials nargs = len(attribute_names) class supertup(tuple): _ _slots_ _ = ( ) # save memory, we don't need per-instance dict def _ _new_ _(cls, *args): if len(args) != nargs: raise TypeError, '%s takes exactly %d arguments (%d given)' % ( typename, nargs, len(args)) return tuple._ _new_ _(cls, args) def _ _repr_ _(self): return '%s(%s)' % (typename, ', '.join(map(repr, self))) # add a few key touches to our new subclass of `tuple' for index, attr_name in enumerate(attribute_names): setattr(supertup, attr_name, property(itemgetter(index))) supertup._ _name_ _ = typename return supertup Discussion You often want to pass data around by means of tuples, which play the role of C's structs, or that of simple records in other languages. Having to remember which numeric index corresponds to which field, and accessing the fields by indexing, is often bothersome. Some Python Standard Library modules, such as time and os, which in old Python versions used to return tuples, have fixed the problem by returning, instead, instances of tuple-like types that let you access the fields by name, as attributes, as well as by index, as items. This recipe shows you how to get the same effect for your code, essentially by automatically building a custom subclass of tuple.Orchestrating the building of a new, customized type can be achieved in several ways; custom metaclasses are often the best approach for such tasks. In this case, however, a simple factory function is quite sufficient, and you should never use more power than you need. Here is how you can use this recipe's superTuple factory function in your code, assuming you have saved this recipe's Solution as a module named supertuple.py somewhere along your Python sys.path: >>> import supertuple >>> Point = supertuple.superTuple('Point', 'x', 'y') >>> Point >>> p = Point(1, 2, 3) # wrong number of fields Traceback (most recent call last): File "", line 1, in ? File "C:\Python24\Lib\site-packages\superTuple.py", line 16, in _ _new_ _ raise TypeError, '%s takes exactly %d arguments (%d given)' % ( TypeError: Point takes exactly 2 arguments (3 given) >>> p = Point(1, 2) # let's do it right this time >>> p Point(1, 2) >>> print p.x, p.y 1 2 Function superTuple's implementation is quite straightforward. To build the new subclass, superTuple uses a class statement, and in that statement's body, it defines three specials: an "empty" _ _slots_ _ (just to save memory, since our supertuple instances don't need any per- instance dictionary anyway); a _ _new_ _ method that checks the number of arguments before delegating to tuple._ _new_ _; and an appropriate _ _repr_ _ method. After the new class object is built, we set into it a property for each named attribute we want. Each such property has only a "getter", since our supertuples, just like tuples themselves, are immutableno setting of fields. Finally, we set the new class' name and return the class object. Each of the getters is easily built by a simple call to the built-in itemgetter from the standard library module operator. Since operator.itemgetter was introduced in Python 2.4, at the very start of our module we ensure we have a suitable itemgetter at hand anyway, even in Python 2.3, by rolling our own if necessary. See Also Library Reference and Python in a Nutshell docs for property, _ _slots_ _, tuple, and special methods _ _new_ _ and _ _repr_ _; (Python 2.4 only) module operator's function itemgetter. Recipe 6.8. Avoiding Boilerplate Accessors for Properties Credit: Yakov Markovitch Problem Your classes use some property instances where either the getter or the setter is just boilerplate code to fetch or set an instance attribute. You would prefer to just specify the attribute name, instead of writing boilerplate code. Solution You need a factory function that catches the cases in which either the getter or the setter argument is a string, and wraps the appropriate argument into a function, then delegates the rest of the work to Python's built-in property: def xproperty(fget, fset, fdel=None, doc=None): if isinstance(fget, str): attr_name = fget def fget(obj): return getattr(obj, attr_name) elif isinstance(fset, str): attr_name = fset def fset(obj, val): setattr(obj, attr_name, val) else: raise TypeError, 'either fget or fset must be a str' return property(fget, fset, fdel, doc) Discussion Python's built-in property is very useful, but it presents one minor annoyance (it may be easier to see as an annoyance for programmers with experience in Delphi). It often happens that you want to have both a setter and a "getter", but only one of them actually needs to execute any significant code; the other one simply needs to read or write an instance attribute. In that case, property still requires two functions as its arguments. One of the functions will then be just "boilerplate code" (i.e., repetitious plumbing code that is boring, and often voluminous, and thus a likely home for bugs). For example, consider: class Lower(object): def _ _init_ _(self, s=''): self.s = s def _getS(self): return self._s def _setS(self, s): self._s = s.lower( ) s = property(_getS, _setS)Method _getS is just boilerplate, yet you have to code it because you need to pass it to property. Using this recipe, you can make your code a little bit simpler, without changing the code's meaning: class Lower(object): def _ _init_ _(self, s=''): self.s = s def _setS(self, s): self._s = s.lower( ) s = xproperty('_s', _setS) The simplification doesn't look like much in one small example, but, applied widely all over your code, it can in fact help quite a bit. The implementation of factory function xproperty in this recipe's Solution is rather rigidly coded: it requires you to pass both fget and fset, and exactly one of them must be a string. No use case requires that both be strings; when neither is a string, or when you want to have just one of the two accessors, you can (and should) use the built-in property directly. It is better, therefore, to have xproperty check that it is being used accurately, considering that such checks remove no useful functionality and impose no substantial performance penalty either. See Also Library Reference and Python in a Nutshell documentation on the built-in property. Recipe 6.9. Making a Fast Copy of an Object Credit: Alex Martelli Problem You need to implement the special method _ _copy_ _ so that your class can cooperate with the copy.copy function. Because the _ _init_ _ method of your specific class happens to be slow, you need to bypass it and get an "empty", uninitialized instance of the class. Solution Here's a solution that works for both new-style and classic classes: def empty_copy(obj): class Empty(obj._ _class_ _): def _ _init_ _(self): pass newcopy = Empty( ) newcopy._ _class_ _ = obj._ _class_ _ return newcopy Your classes can use this function to implement _ _copy_ _ as follows: class YourClass(object): def _ _init_ _(self): assume there's a lot of work here def _ _copy_ _(self): newcopy = empty_copy(self) copy some relevant subset of self's attributes to newcopy return newcopy Here's a usage example: if _ _name_ _ == '_ _main_ _': import copy y = YourClass( ) # This, of course, does run _ _init_ _ print y z = copy.copy(y) # ...but this doesn't print z Discussion As covered in Recipe 4.1, Python doesn't implicitly copy your objects when you assign them, which is a great thing because it gives fast, flexible, and uniform semantics. When you need a copy, you explicitly ask for it, often with the copy.copy function, which knows how to copy built- in types, has reasonable defaults for your own objects, and lets you customize the copying process by defining a special method _ _copy_ _ in your own classes. If you want instances of a class to be noncopyable, you can define _ _copy_ _ and raise a TypeError there. In most cases,you can just let copy.copy's default mechanisms work, and you get free clonability for most of your classes. This is quite a bit nicer than languages that force you to implement a specific clone method for every class whose instances you want to be clonable. A _ _copy_ _ method often needs to start with an "empty" instance of the class in question (e.g., self), bypassing _ _init_ _ when that is a costly operation. The simplest general way to do this is to use the ability that Python gives you to change an instance's class on the fly: create a new object in a local empty class, then set the new object's _ _class_ _ attribute, as the recipe's code shows. Inheriting class Empty from obj._ _class_ _ is redundant (but quite innocuous) for old-style (classic) classes, but that inheritance makes the recipe compatible with all kinds of objects of classic or new-style classes (including built-in and extension types). Once you choose to inherit from obj's class, you must override _ _init_ _ in class Empty, or else the whole purpose of the recipe is defeated. The override means that the _ _init_ _ method of obj's class won't execute, since Python, fortunately, does not automatically execute ancestor classes' initializers. Once you have an "empty" object of the required class, you typically need to copy a subset of self's attributes. When you need all of the attributes, you're better off not defining _ _copy_ _ explicitly, since copying all instance attributes is exactly copy.copy's default behavior. Unless, of course, you need to do a little bit more than just copying instance attributes; in this case, these two alternative techniques to copy all attributes are both quite acceptable: newcopy._ _dict_ _.update(self._ _dict_ _) newcopy._ _dict_ _ = dict(self._ _dict_ _) An instance of a new-style class doesn't necessarily keep all of its state in _ _dict_ _, so you may need to do some class-specific state copying in such cases. Alternatives based on the new standard module can't be made transparent across classic and new-style classes, and neither can the _ _new_ _ static method that generates an empty instancethe latter is only defined in new-style classes, not classic ones. Fortunately, this recipe obviates any such issues. A good alternative to implementing _ _copy_ _ is often to implement the methods _ _getstate_ _ and _ _setstate_ _ instead: these special methods define your object's state very explicitly and intrinsically bypass _ _init_ _. Moreover, they also support serialization (i.e., pickling) of your class instances: see Recipe 7.4 for more information about these methods. So far we have been discussing shallow copies, which is what you want most of the time. With a shallow copy, your object is copied, but objects it refers to (attributes or items) are not, so the newly copied object and the original object refer to the same items or attributes objectsa fast and lightweight operation. A deep copy is a heavyweight operation, potentially duplicating a large graph of objects that refer to each other. You get a deep copy by calling copy.deepcopy on an object. If you need to customize the way in which instances of your class are deep-copied, you can define the special method _ _deepcopy_ _: class YourClass(object): ... def _ _deepcopy_ _(self, memo): newcopy = empty_copy(self) # use copy.deepcopy(self.x, memo) to get deep copies of elements # in the relevant subset of self's attributes, to set in newcopy return newcopy If you choose to implement _ _deepcopy_ _, remember to respect the memoization protocol that is specified in the Python documentation for standard module copyget deep copies of all the attributes or items that are needed by calling copy.deepcopy with a second argument, the same memo dictionary that is passed to the _ _deepcopy_ _ method. Again, implementing _ _getstate_ _ and _ _setstate_ _ is often a good alternative, since these methods can alsosupport deep copying: Python takes care of deeply copying the "state" object that _ _getstate_ _ returns, before passing it to the _ _setstate_ _ method of a new, empty instance. See Recipe 7.4 for more information about these special methods. See Also Recipe 4.1 about shallow and deep copies; Recipe 7.4 about _ _getstate_ _ and _ _setstate_ _; the Library Reference and Python in a Nutshell sections on the copy module. Recipe 6.10. Keeping References to Bound Methods Without Inhibiting Garbage Collection Credit: Joseph A. Knapka, Frédéric Jolliton, Nicodemus Problem You want to hold references to bound methods, while still allowing the associated object to be garbage-collected. Solution Weak references (i.e., references that indicate an object as long as that object is alive but don't keep that object alive if there are no other, normal references to it) are an important tool in some advanced programming situations. The weakref module in the Python Standard Library lets you use weak references. However, weakref's functionality cannot directly be used for bound methods unless you take some precautions. To allow an object to be garbage-collected despite outstanding references to its bound methods, you need some wrappers. Put the following code in a file named weakmethod.py in some directory on your Python sys.path: import weakref, new class ref(object): """ Wraps any callable, most importantly a bound method, in a way that allows a bound method's object to be GC'ed, while providing the same interface as a normal weak reference. """ def _ _init_ _(self, fn): try: # try getting object, function, and class o, f, c = fn.im_self, fn.im_func, fn.im_class except AttributeError: # It's not a bound method self._obj = None self._func = fn self._clas = None else: # It is a bound method if o is None: self._obj = None # ...actually UN-bound else: self._obj = weakref.ref(o) # ...really bound self._func = f self._clas = c def _ _call_ _(self): if self.obj is None: return self._func elif self._obj( ) is None: return None return new.instancemethod(self._func, self.obj( ), self._clas) Discussion A normal bound method holds a strong reference to the bound method's object. That means that the object can't be garbage-collected until the bound method is disposed of:>>> class C(object): ... def f(self): ... print "Hello" ... def _ _del_ _(self): ... print "C dying" ... >>> c = C( ) >>> cf = c.f >>> del c # c continues to wander about with glazed eyes... >>> del cf # ...until we stake its bound method, only then it goes away: C dying This behavior is most often handy, but sometimes it's not what you want. For example, if you're implementing an event-dispatch system, it might not be desirable for the mere presence of an event handler (i.e., a bound method) to prevent the associated object from being reclaimed. The instinctive idea should then be to use weak references. However, a normal weakref.ref to a bound method doesn't quite work the way one might expect, because bound methods are first- class objects. Weak references to bound methods are dead-on-arrivalthat is, they always return None when dereferenced, unless another strong reference to the same bound-method object exists. For example, the following code, based on the weakref module from the Python Standard Library, doesn't print "Hello" but raises an exception instead: >>> import weakref >>> c = C( ) >>> cf = weakref.ref(c.f) >>> cf # Oops, better try the lightning again, Igor... >>> cf( )( ) Traceback (most recent call last): File "", line 1, in ? TypeError: object of type 'None' is not callable On the other hand, the class ref in the weakmethod module shown in this recipe allows you to have weak references to bound methods in a useful way: >>> import weakmethod >>> cf = weakmethod.ref(c.f) >>> cf( )( ) # It LIVES! Bwahahahaha! Hello >>> del c # ...and it dies C dying >>> print cf( ) None Calling the weakmethod.ref instance, which refers to a bound method, has the same semantics as calling a weakref.ref instance that refers to, say, a function object: if the referent has died, it returns None; otherwise, it returns the referent. Actually, in this case, it returns a freshly minted new.instancemethod (holding a strong reference to the objectso, be sure not to hold on to that, unless you do want to keep the object alive for a while!). Note that the recipe is carefully coded so you can wrap into a ref instance any callable you want, be it a method (bound or unbound), a function, whatever; the weak references semantics, however, are provided only when you're wrapping a bound method; otherwise, ref acts as a normal (strong) reference, holding the callable alive. This basically lets you use ref for wrapping arbitrary callables without needing to check for special cases.If you want semantics closer to that of a weakref.proxy, they're easy to implement, for example by subclassing the ref class given in this recipe. When you call a proxy, the proxy calls the referent with the same arguments. If the referent's object no longer lives, then weakref.ReferenceError gets raised instead. Here's an implementation of such a proxy class: class proxy(ref): def _ _call_ _(self, *args, **kwargs): func = ref._ _call_ _(self) if func is None: raise weakref.ReferenceError('referent object is dead') else: return func(*args, **kwargs) def _ _eq_ _(self, other): if type(other) != type(self): return False return ref._ _call_ _(self) == ref._ _call_ _(other) See Also The Library Reference and Python in a Nutshell sections on the weakref and new modules and on bound-method objects. Recipe 6.11. Implementing a Ring Buffer Credit: Sébastien Keim, Paul Moore, Steve Alexander, Raymond Hettinger Problem You want to define a buffer with a fixed size, so that, when it fills up, adding another element overwrites the first (oldest) one. This kind of data structure is particularly useful for storing log and history information. Solution This recipe changes the buffer object's class on the fly, from a nonfull buffer class to a full buffer class, when the buffer fills up: class RingBuffer(object): """ class that implements a not-yet-full buffer """ def _ _init_ _(self, size_max): self.max = size_max self.data = [ ] class _ _Full(object): """ class that implements a full buffer """ def append(self, x): """ Append an element overwriting the oldest one. """ self.data[self.cur] = x self.cur = (self.cur+1) % self.max def tolist(self): """ return list of elements in correct order. """ return self.data[self.cur:] + self.data[:self.cur] def append(self, x): """ append an element at the end of the buffer. """ self.data.append(x) if len(self.data) == self.max: self.cur = 0 # Permanently change self's class from non-full to full self._ _class_ _ = _ _Full def tolist(self): """ Return a list of elements from the oldest to the newest. """ return self.data # sample usage if _ _name_ _ == '_ _main_ _': x = RingBuffer(5) x.append(1); x.append(2); x.append(3); x.append(4) print x._ _class_ _, x.tolist( ) x.append(5) print x._ _class_ _, x.tolist( ) x.append(6) print x.data, x.tolist( ) x.append(7); x.append(8); x.append(9); x.append(10) print x.data, x.tolist( )Discussion A ring buffer is a buffer with a fixed size. When it fills up, adding another element overwrites the oldest one that was still being kept. It's particularly useful for the storage of log and history information. Python has no direct support for this kind of structure, but it's easy to construct one. The implementation in this recipe is optimized for element insertion. The notable design choice in the implementation is that, since these objects undergo a nonreversible state transition at some point in their lifetimesfrom nonfull buffer to full buffer (and behavior changes at that point)I modeled that by changing self._ _class_ _. This works just as well for classic classes as for new-style ones, as long as the old and new classes of the object have the same slots (e.g., it works fine for two new-style classes that have no slots at all, such as RingBuffer and _ _Full in this recipe). Note that, differently from other languages, the fact that class _ _Full is implemented inside class RingBuffer does not imply any special relationship between these classes; that's a good thing, too, because no such relationship is necessary. Changing the class of an instance may be strange in many languages, but it is an excellent Pythonic alternative to other ways of representing occasional, massive, irreversible, and discrete changes of state that vastly affect behavior, as in this recipe. Fortunately, Python supports it for all kinds of classes. Ring buffers (i.e., bounded queues, and other names) are quite a useful idea, but the inefficiency of testing whether the ring is full, and if so, doing something different, is a nuisance. The nuisance is particularly undesirable in a language like Python, where there's no difficultyother than the massive memory cost involvedin allowing the list to grow without bounds. So, ring buffers end up being underused in spite of their potential. The idea of assigning to _ _class_ _ to switch behaviors when the ring gets full is the key to this recipe's efficiency: such class switching is a one-off operation, so it doesn't make the steady-state cases any less efficient. Alternatively, we might switch just two methods, rather than the whole class, of a ring buffer instance that becomes full: class RingBuffer(object): def _ _init_ _(self,size_max): self.max = size_max self.data = [ ] def _full_append(self, x): self.data[self.cur] = x self.cur = (self.cur+1) % self.max def _full_get(self): return self.data[self.cur:]+self.data[:self.cur] def append(self, x): self.data.append(x) if len(self.data) == self.max: self.cur = 0 # Permanently change self's methods from non-full to full self.append = self._full_append self.tolist = self._full_get def tolist(self): return self.data This method-switching approach is essentially equivalent to the class-switching one in the recipe's solution, albeit through rather different mechanisms. The best approach is probably to use class switching when all methods must be switched in bulk and method switching only when you need finer granularity of behavior change. Class switching is the only approach that works if you need to switch any special methods in a new-style class, since intrinsic lookup of special methods during various operations happens on the class, not on the instance (classic classes differ from new-style ones in this aspect).You can use many other ways to implement a ring buffer. In Python 2.4, in particular, you should consider subclassing the new type collections.deque, which supplies a "double-ended queue", allowing equally effective additions and deletions from either end: from collections import deque class RingBuffer(deque): def _ _init_ _(self, size_max): deque._ _init_ _(self) self.size_max = size_max def append(self, datum): deque.append(self, datum) if len(self) > self.size_max: self.popleft( ) def tolist(self): return list(self) or, to avoid the if statement when at steady state, you can mix this idea with the idea of switching a method: from collections import deque class RingBuffer(deque): def _ _init_ _(self, size_max): deque._ _init_ _(self) self.size_max = size_max def _full_append(self, datum): deque.append(self, datum) self.popleft( ) def append(self, datum): deque.append(self, datum) if len(self) == self.size_max: self.append = self._full_append def tolist(self): return list(self) With this latest implementation, we need to switch only the append method (the tolist method remains the same), so method switching appears to be more appropriate than class switching. See Also The Reference Manual and Python in a Nutshell sections on the standard type hierarchy and classic and new-style object models; Python 2.4 Library Reference on module collections. Recipe 6.12. Checking an Instance for Any State Changes Credit: David Hughes Problem You need to check whether any changes to an instance's state have occurred to selectively save instances that have been modified since the last "save" operation. Solution An effective solution is a mixin classa class you can multiply inherit from and that is able to take snapshots of an instance's state and compare the instance's current state with the last snapshot to determine whether or not the instance has been modified: import copy class ChangeCheckerMixin(object): containerItems = {dict: dict.iteritems, list: enumerate} immutable = False def snapshot(self): ''' create a "snapshot" of self's state -- like a shallow copy, but recursing over container types (not over general instances: instances must keep track of their own changes if needed). ''' if self.immutable: return self._snapshot = self._copy_container(self._ _dict_ _) def makeImmutable(self): ''' the instance state can't change any more, set .immutable ''' self.immutable = True try: del self._snapshot except AttributeError: pass def _copy_container(self, container): ''' semi-shallow copy, recursing on container types only ''' new_container = copy.copy(container) for k, v in self.containerItems[type(new_container)](new_container): if type(v) in self.containerItems: new_container[k] = self._copy_container(v) elif hasattr(v, 'snapshot'): v.snapshot( ) return new_container def isChanged(self): ''' True if self's state is changed since the last snapshot ''' if self.immutable: return False # remove snapshot from self._ _dict_ _, put it back at the end snap = self._ _dict_ _.pop('_snapshot', None) if snap is None: return True try: return self._checkContainer(self._ _dict_ _, snap) finally: self._snapshot = snap def _checkContainer(self, container, snapshot): ''' return True if the container and its snapshot differ ''' if len(container) != len(snapshot): return True for k, v in self.containerItems[type(container)](container): try: ov = snapshot[k] except LookupError: return True if self._checkItem(v, ov): return True return False def _checkItem(self, newitem, olditem): ''' compare newitem and olditem. If they are containers, call self._checkContainer recursively. If they're an instance with an 'isChanged' method, delegate to that method. Otherwise, return True if the items differ. ''' if type(newitem) != type(olditem): return True if type(newitem) in self.containerItems: return self._checkContainer(newitem, olditem) if newitem is olditem: method_isChanged = getattr(newitem, 'isChanged', None) if method_isChanged is None: return False return method_isChanged( ) return newitem != olditem Discussion I often need change-checking functionality in my applications. For example, when a user closes the last GUI window over a certain document, I need to check whether the document was changed since the last "save" operation; if it was, then I need to pop up a small window to give the user a choice between saving the document, losing the latest changes, or canceling the window-closing operation. The class ChangeCheckerMixin, which this recipe describes, satisfies this need. The idea is to multiply derive all of your data classes, meaning all classes that hold data the user views and may change, from ChangeCheckerMixin (as well as from any other bases they need). When the data has just been loaded from or saved to persistent storage, call method snapshot on the top- level, document data class instance. This call takes a "snapshot" of the current state, basically a shallow copy of the object but with recursion over containers, and calls the snapshot methods on any contained instance that has such a method. Any time afterward, you can call method isChanged on any data class instance to check whether the instance state was changed since the time of its last snapshot. As container types, ChangeCheckerMixin, as presented, considers only list and dict. If you also use other types as containers, you just need to add them appropriately to the containerItems dictionary. That dictionary must map each container type to a function callable on an instance of that type to get an iterator on indices and values (with indices usable to index the container). Container type instances must also support being shallowly copied with standard library Python function copy.copy. For example, to add Python 2.4's collections.deque as a container to a subclass of ChangeCheckerMixin, you can code: import collectionsclass CCM_with_deque(ChangeCheckerMixin): containerItems = dict(ChangeCheckerMixin.containerItems) containerItems[collections.deque] = enumerate since collections.deque can be "walked over" with enumerate, just like list can. Here is a toy example of use for ChangeChecherMixin: if _ _name_ _ == '_ _main_ _': class eg(ChangeCheckerMixin): def _ _init_ _(self, *a, **k): self.L = list(*a, **k) def _ _str_ _(self): return 'eg(%s)' % str(self.L) def _ _getattr_ _(self, a): return getattr(self.L, a) x = eg('ciao') print 'x =', x, 'is changed =', x.isChanged( ) # emits: x = eg(['c', 'i', 'a', 'o']) is changed = True # now, assume x gets saved, then...: x.snapshot( ) print 'x =', x, 'is changed =', x.isChanged( ) # emits: x = eg(['c', 'i', 'a', 'o']) is changed = False # now we change x...: x.append('x') print 'x =', x, 'is changed =', x.isChanged( ) # emits: x = eg(['c', 'i', 'a', 'o', 'x']) is changed = True In class eg we only subclass ChanceCheckerMixin because we need no other bases. In particular, we cannot usefully subclass list because the change-checking functionality works only on state that is kept in an instance's dictionary; so, we must hold a list object in our instance's dictionary, and delegate to it as needed (in this toy example, we delegate all nonspecial methods, automatically, via _ _getattr_ _). With this precaution, we see that the isChanged method correctly reflects the crucial tidbitwhether the instance's state has been changed since the last call to snapshot on the instance. An implicit assumption of this recipe is that your application's data class instances are organized in a hierarchical fashion. The tired old (but still valid) example is an invoice containing header data and detail lines. Each instance of the details data class could contain other instances, such as product details, which may not be modifiable in the current activity but are probably modifiable elsewhere. This is the reason for the immutable attribute and the makeImmutable method: when the attribute is set by calling the method, any outstanding snapshot for the instance is dropped to save memory, and further calls to either snapshot or isChanged can return very rapidly. If your data does not lend itself to such hierarchical structuring, you may have to take full deep copies, or even "snapshot" a document instance by taking a full pickle of it, and check for changes by comparing the new pickle with the last one previously taken. That may be all right on very fast machines, or when the amount of data you're handling is rather modest. In my tests, however, it shows up as being unacceptably slow for substantial amounts of data on more ordinary machines. This recipe, when your data organization is suitable for its application, can offer better performance. If some of your data classes also contain data that is automatically computed or, for other reasons, does not need to be saved, store such data in instances of subordinate classes (which do not inherit from ChangeCheckerMixin), rather than either holding the data as attributes or storing it in ordinary containers such as lists and dictionaries. See AlsoLibrary Reference and Python in a Nutshell documentation on multiple inheritance, the iteritems method of dictionaries, and built-in functions enumerate, isinstance, and hasattr. Recipe 6.13. Checking Whether an Object Has Necessary Attributes Credit: Alex Martelli Problem You need to check whether an object has certain necessary attributes before performing state- altering operations. However, you want to avoid type-testing because you know it interferes with polymorphism. Solution In Python, you normally just try performing whatever operations you need to perform. For example, here's the simplest, no-checks code for doing a certain sequence of manipulations on a list argument: def munge1(alist): alist.append(23) alist.extend(range(5)) alist.append(42) alist[4] = alist[3] alist.extend(range(2)) If alist is missing any of the methods you're calling (explicitly, such as append and extend; or implicitly, such as the calls to _ _getitem_ _ and _ _setitem_ _ implied by the assignment statement alist[4] = alist[3]), the attempt to access and call a missing method raises an exception. Function munge1 makes no attempt to catch the exception, so the execution of munge1 terminates, and the exception propagates to the caller of munge1. The caller may choose to catch the exception and deal with it, or terminate execution and let the exception propagate further back along the chain of calls, as appropriate. This approach is usually just fine, but problems may occasionally occur. Suppose, for example, that the alist object has an append method but not an extend method. In this peculiar case, the munge1 function partially alters alist before an exception is raised. Such partial alterations are generally not cleanly undoable; depending on your application, they can sometimes be a bother. To forestall the "partial alterations" problem, the first approach that comes to mind is to check the type of alist. Such a naive "Look Before You Leap" (LBYL) approach may look safer than doing no checks at all, but LBYL has a serious defect: it loses polymorphism! The worst approach of all is checking for equality of types: def munge2(alist): if type(alist) is list: # a very bad idea munge1(alist) else: raise TypeError, "expected list, got %s" % type(alist) This even fails, without any good reason, when alist is an instance of a subclass of list. You can at least remove that huge defect by using isinstance instead:def munge3(alist): if isinstance(alist, list): munge1(alist) else: raise TypeError, "expected list, got %s" % type(alist) However, munge3 still fails, needlessly, when alist is an instance of a type or class that mimics list but doesn't inherit from it. In other words, such type-checking sacrifices one of Python's great strengths: signature-based polymorphism. For example, you cannot pass to munge3 an instance of Python 2.4's collections.deque, which is a real pity because such a deque does supply all needed functionality and indeed can be passed to the original munge1 and work just fine. Probably a zillion sequence types are out there that, like deque, are quite acceptable to munge1 but not to munge3. Type-checking, even with isinstance, exacts an enormous price. A far better solution is accurate LBYL, which is both safe and fully polymorphic: def munge4(alist): # Extract all bound methods you need (get immediate exception, # without partial alteration, if any needed method is missing): append = alist.append extend = alist.extend # Check operations, such as indexing, to get an exception ASAP # if signature compatibility is missing: try: alist[0] = alist[0] except IndexError: pass # An empty alist is okay # Operate: no exceptions are expected from this point onwards append(23) extend(range(5)) append(42) alist[4] = alist[3] extend(range(2)) Discussion Python functions are naturally polymorphic on their arguments because they essentially depend on the methods and behaviors of the arguments, not on the arguments' types. If you check the types of arguments, you sacrifice this precious polymorphism, so, don't! However, you may perform a few early checks to obtain some extra safety (particularly against partial alterations) without substantial costs.What Is Polymorphism? Polymorphism (from Greek roots meaning "many shapes") is the ability of code to deal with objects of different types in ways that are appropriate to each applicable type. Unfortunately, this useful term has been overloaded with all sorts of implications, to the point that many people think it's somehow connected with such concepts as overloading (specifying different functions depending on call-time signatures) or subtyping (i.e., subclassing), which it most definitely isn't. Subclassing is often a useful implementation technique, but it's not a necessary condition for polymorphism. Overloading is right out: Python just doesn't let multiple objects with the same name live at the same time in the same scope, so you can't have several functions or methods with the same name and scope, distinguished only by their signaturesa minor annoyance, at worst: just rename those functions or methods so that their name suffices to distinguish them. Python's functions are polymorphic (unless you take specific steps to break this very useful feature) because they just call methods on their arguments (explicitly or implicitly by performing operations such as arithmetic and indexing): as long as the arguments supply the needed methods, callable with the needed signatures, and those calls perform the appropriate behavior, everything just works. The normal Pythonic way of life can be described as the Easier to Ask Forgiveness than Permission (EAFP) approach: just try to perform whatever operations you need, and either handle or propagate any exceptions that may result. It usually works great. The only real problem that occasionally arises is "partial alteration": when you need to perform several operations on an object, just trying to do them all in natural order could result in some of them succeeding, and partially altering the object, before an exception is raised. For example, suppose that munge1, as shown at the start of this recipe's Solution, is called with an actual argument value for alist that has an append method but lacks extend. In this case, alist is altered by the first call to append; but then, the attempt to obtain and call extend raises an exception, leaving alist's state partially altered, a situation that may be hard to recover from. Sometimes, a sequence of operations should ideally be atomic: either all of the alterations happen, and everything is fine, or none of them do, and an exception gets raised. You can get closer to ideal atomicity by switching to the LBYL approach, but in an accurate, careful way. Extract all bound methods you'll need, then noninvasively test the necessary operations (such as indexing on both sides of the assignment operator). Move on to actually changing the object state only if all of this succeeds. From that point onward, it's far less likely (although not impossible) that exceptions will occur in midstream, leaving state partially altered. You could not reach 100% safety even with the strictest type-checking, after all: for example, you might run out of memory just smack in the middle of your operations. So, with or without type-checking, you don't really ever guarantee atomicityyou just approach asymptotically to that desirable property. Accurate LBYL generally offers a good trade-off in comparison to EAFP, assuming we need safeguards against partial alterations. The extra complication is modest, and the slowdown due to the checks is typically compensated by the extra speed gained by using bound methods through local names rather than explicit attribute access (at least if the operations include loops, which is often the case). It's important to avoid overdoing the checks, and the assert statement can help with that. For example, you can add such checks as assert callable(append) to munge4. In this case, the compiler removes the assert entirely when you run the program with optimization (i.e., with flags -O or -OO passed to the python command), while performing the checks when the program is run for testing and debugging (i.e., without the optimization flags).See Also Language Reference and Python in a Nutshell about assert and the meaning of the -O and -OO command-line arguments; Library Reference and Python in a Nutshell about sequence types, and lists in particular. Recipe 6.14. Implementing the State Design Pattern Credit: Elmar Bschorer Problem An object in your program can switch among several "states", and the object's behavior must change along with the object's state. Solution The key idea of the State Design Pattern is to objectify the "state" (with its several behaviors) into a class instance (with its several methods). In Python, you don't have to build an abstract class to represent the interface that is common to the various states: just write the classes for the "state"s themselves. For example: class TraceNormal(object): ' state for normal level of verbosity ' def startMessage(self): self.nstr = self.characters = 0 def emitString(self, s): self.nstr += 1 self.characters += len(s) def endMessage(self): print '%d characters in %d strings' % (self.characters, self.nstr) class TraceChatty(object): ' state for high level of verbosity ' def startMessage(self): self.msg = [ ] def emitString(self, s): self.msg.append(repr(s)) def endMessage(self): print 'Message: ', ', '.join(self.msg) class TraceQuiet(object): ' state for zero level of verbosity ' def startMessage(self): pass def emitString(self, s): pass def endMessage(self): pass class Tracer(object): def _ _init_ _(self, state): self.state = state def setState(self, state): self.state = state def emitStrings(self, strings): self.state.startMessage( ) for s in strings: self.state.emitString(s) self.state.endMessage( ) if _ _name_ _ == '_ _main_ _': t = Tracer(TraceNormal( )) t.emitStrings('some example strings here'.split( )) # emits: 21 characters in 4 strings t.setState(TraceQuiet( )) t.emitStrings('some example strings here'.split( )) # emits nothing t.setState(TraceChatty( )) t.emitStrings('some example strings here'.split( )) # emits: Message: 'some', 'example', 'strings', 'here' Discussion With the State Design Pattern, you can "factor out" a number of related behaviors of an object (and possibly some data connected with these behaviors) into an auxiliary state object, to which the main object delegates these behaviors as needed, through calls to methods of the "state" object. In Python terms, this design pattern is related to the idioms of rebinding an object's whole _ _class_ _, as shown in Recipe 6.11, and rebinding just certain methods (shown in Recipe 2.14). This design pattern, in a sense, lies in between those Python idioms: you group a set of related behaviors, rather than switching either all behavior, by changing the object's whole _ _class_ _, or each method on its own, without grouping. With relation to the classic design pattern terminology, this recipe presents a pattern that falls somewhere between the classic State Design Pattern and the classic Strategy Design Pattern. This State Design Pattern has some extra oomph, compared to the related Pythonic idioms, because an appropriate amount of data can live together with the behaviors you're delegatingexactly as much, or as little, as needed to support each specific behavior. In the examples given in this recipe's Solution, for example, the different state objects differ greatly in the kind and amount of data they need: none at all for class TraceQuiet, just a couple of numbers for TraceNormal, a whole list of strings for TraceChatty. These responsibilities are usefully delegated from the main object to each specific "state object". In some cases, although not in the specific examples shown in this recipe, state objects may need to cooperate more closely with the main object, by calling main object methods or accessing main object attributes in certain circumstances. To allow this, the main object can pass as an argument either self or some bound method of self to methods of the "state" objects. For example, suppose that the functionality in this recipe's Solution needs to be extended, in that the main object must keep track of how many lines have been emitted by messages it has sent. Tracer._ _init_ _ will have to add one per-instance initialization self.lines = 0, and the signature of the "state" object's endMessage methods will have to be extended to def endMessage(self, tracer):. The implementation of endMessage in class TraceQuiet will just ignore the tracer argument, since it doesn't actually emit any lines; the implementations in the other two classes will each add a statement tracer.lines += 1, since each of them emits one line per message. As you see, the kind of closer coupling implied by this kind of extra functionality need not be particularly problematic. In particular, the key feature of the classic State Design Pattern, that state objects are the ones that handle state switching (while, in the Strategy Design Pattern, the switching comes from the outside), is just not enough of a big deal in Python to warrant considering the two design patterns as separate. See Also See http://exciton.cs.rice.edu/JavaResources/DesignPatterns/ for good coverage of the classic design patterns, albeit in a Java context. Recipe 6.15. Implementing the "Singleton" Design Pattern Credit: Jürgen Hermann Problem You want to make sure that only one instance of a class is ever created. Solution The _ _new_ _ staticmethod makes the task very simple: class Singleton(object): """ A Pythonic Singleton """ def _ _new_ _(cls, *args, **kwargs): if '_inst' not in vars(cls): cls._inst = type._ _new_ _(cls, *args, **kwargs) return cls._inst Just have your class inherit from Singleton, and don't override _ _new_ _. Then, all calls to that class (normally creations of new instances) return the same instance. (The instance is created once, on the first such call to each given subclass of Singleton during each run of your program.) Discussion This recipe shows the one obvious way to implement the "Singleton" Design Pattern in Python (see E. Gamma, et al., Design Patterns: Elements of Reusable Object-Oriented Software, Addison-Wesley). A Singleton is a class that makes sure only one instance of it is ever created. Typically, such a class is used to manage resources that by their nature can exist only once. See Recipe 6.16 for other considerations about, and alternatives to, the "Singleton" design pattern in Python. We can complete the module with the usual self-test idiom and show this behavior: if _ _name_ _ == '_ _main_ _': class SingleSpam(Singleton): def _ _init_ _(self, s): self.s = s def _ _str_ _(self): return self.s s1 = SingleSpam('spam') print id(s1), s1.spam( ) s2 = SingleSpam('eggs') print id(s2), s2.spam( ) When we run this module as a script, we get something like the following output (the exact value of id does vary, of course): 8172684 spam8172684 spam The 'eggs' parameter passed when trying to instantiate s2 has been ignored, of coursethat's part of the price you pay for having a Singleton! One issue with Singleton in general is subclassability. The way class Singleton is coded in this recipe, each descendant subclass, direct or indirect, will get a separate instance. Literally speaking, this violates the constraint of only one instance per class, depending on what one exactly means by it: class Foo(Singleton): pass class Bar(Foo): pass f = Foo( ); b = Bar( ) print f is b, isinstance(f, Foo), isinstance(b, Foo) # emits False True True f and b are separate instances, yet, according to the built-in function isinstance, they are both instances of Foo because isinstance applies the IS-A rule of OOP: an instance of a subclass IS- An instance of the base class too. On the other hand, if we took pains to return f again when b is being instantiated by calling Bar, we'd be violating the normal assumption that calling class Bar gives us an instance of class Bar, not an instance of a random superclass of Bar that just happens to have been instantiated earlier in the course of a run of the program. In practice, subclassability of "Singleton"s is rather a headache, without any obvious solution. If this issue is important to you, the alternative Borg idiom, explained next in Recipe 6.16 may provide a better approach. See Also Recipe 6.16; E. Gamma, R. Helm, R. Johnson, J. Vlissides, Design Patterns: Elements of Reusable Object-Oriented Software (Addison-Wesley). Recipe 6.16. Avoiding the "Singleton" Design Pattern with the Borg Idiom Credit: Alex Martelli, Alex A. Naanou Problem You want to make sure that only one instance of a class is ever created: you don't care about the id of the resulting instances, just about their state and behavior, and you need to ensure subclassability. Solution Application needs (forces) related to the "Singleton" Design Pattern can be met by allowing multiple instances to be created while ensuring that all instances share state and behavior. This is more flexible than fiddling with instance creation. Have your class inherit from the following Borg class: class Borg(object): _shared_state = { } def _ _new_ _(cls, *a, **k): obj = object._ _new_ _(cls, *a, **k) obj._ _dict_ _ = cls._shared_state return obj If you override _ _new_ _ in your class (very few classes need to do that), just remember to use Borg._ _new_ _, rather than object._ _new_ _, within your override. If you want instances of your class to share state among themselves, but not with instances of other subclasses of Borg, make sure that your class has, at class scope, the "state"ment: _shared_state = { } With this "data override", your class doesn't inherit the _shared_state attribute from Borg but rather gets its own. It is to enable this "data override" that Borg's _ _new_ _ uses cls._shared_state instead of Borg._shared_state. Discussion Borg in action Here's a typical example of Borg use: if _ _name_ _ == '_ _main_ _': class Example(Borg): name = None def _ _init_ _(self, name=None): if name is not None: self.name = name def _ _str_ _(self): return 'name->%s' % self.name a = Example('Lara') b = Example( ) # instantiating b shares self.name with a print a, b c = Example('John Malkovich') # making c changes self.name of a & b too print a, b, c b.name = 'Seven' # setting b.name changes name of a & c too print a, b, c When running this module as a main script, the output is: name->Lara name->Lara name->John Malkovich name->John Malkovich name->John Malkovich name->Seven name->Seven name->Seven All instances of Example share state, so any setting of the name attribute of any instance, either in _ _init_ _ or directly, affects all instances equally. However, note that the instance's ids differ; therefore, since we have not defined special methods _ _eq_ _ and _ _hash_ _, each instance can work as a distinct key in a dictionary. Thus, if we continue our sample code as follows: adict = { } j = 0 for i in a, b, c: adict[i] = j j = j + 1 for i in a, b, c: print i, adict[i] the output is: name->Seven 0 name->Seven 1 name->Seven 2 If this behavior is not what you want, add _ _eq_ _ and _ _hash_ _ methods to the Example class or the Borg superclass. Having these methods might better simulate the existence of a single instance, depending on your exact needs. For example, here's a version of Borg with these special methods added: class Borg(object): _shared_state = { } def _ _new_ _(cls, *a, **k): obj = object._ _new_ _(cls, *a, **k) obj._ _dict_ _ = cls._shared_state return obj def _ _hash_ _(self): return 9 # any arbitrary constant integer def _ _eq_ _(self, other): try: return self._ _dict_ _ is other._ _dict_ _ except AttributeError: return False With this enriched version of Borg, the example's output changes to: name->Seven 2 name->Seven 2name->Seven 2 Borg, Singleton, or neither? The Singleton Design Pattern has a catchy name, but unfortunately it also has the wrong focus for most purposes: it focuses on object identity, rather than on object state and behavior. The Borg design nonpattern makes all instances share state instead, and Python makes implementing this idea a snap. In most cases in which you might think of using Singleton or Borg, you don't really need either of them. Just write a Python module, with functions and module-global variables, instead of defining a class, with methods and per-instance attributes. You need to use a class only if you must be able to inherit from it, or if you need to take advantage of the class' ability to define special methods. (See Recipe 6.2 for a way to combine some of the advantages of classes and modules.) Even when you do need a class, it's usually unnecessary to include in the class itself any code to enforce the idea that one can't make multiple instances of it; other, simpler idioms are generally preferable. For example: class froober(object): def _ _init_ _(self): etc, etc froober = froober( ) Now froober is by nature the only instance of its own class, since name 'froober' has been rebound to mean the instance, not the class. Of course, one might call froober._ _class_ _( ), but it's not sensible to spend much energy taking precautions against deliberate abuse of your design intentions. Any obstacles you put in the way of such abuse, somebody else can bypass. Taking precautions against accidental misuse is way plenty. If the very simple idiom shown in this latest snippet is sufficient for your needs, use it, and forget about Singleton and Borg. Remember: do the simplest thing that could possibly work. On rare occasions, though, an idiom as simple as this one cannot work, and then you do need more. The Singleton Design Pattern (described previously in Recipe 6.15) is all about ensuring that just one instance of a certain class is ever created. In my experience, Singleton is generally not the best solution to the problems it tries to solve, producing different kinds of issues in various object models. We typically want to let as many instances be created as necessary, but all with shared state. Who cares about identity? It's state (and behavior) we care about. The alternate pattern based on sharing state, in order to solve roughly the same problems as Singleton does, has also been called Monostate. Incidentally, I like to call Singleton "Highlander" because there can be only one. In Python, you can implement the Monostate Design Pattern in many ways, but the Borg design nonpattern is often best. Simplicity is Borg's greatest strength. Since the _ _dict_ _ of any instance can be rebound, Borg in its _ _new_ _ rebinds the _ _dict_ _ of each of its instances to a class-attribute dictionary. Now, any reference or binding of an instance attribute will affect all instances equally. I thank David Ascher for suggesting the appropriate name Borg for this nonpattern. Borg is a nonpattern because it had no known uses at the time of its first publication (although several uses are now known): two or more known uses are part of the prerequisites for being a design pattern. See the detailed discussion at http://www.aleax.it/5ep.html. An excellent article by Robert Martin about Singleton and Monostate can be found at http://www.objectmentor.com/resources/articles/SingletonAndMonostate.pdf. Note that most of the disadvantages that Martin attributes to Monostate are really due to the limitations of the languages that Martin is considering, such as C++ and Java, and just disappear when using Borg in Python. For example, Martin indicates, as Monostate's first and main disadvantage, that "A non-Monostate class cannot be converted into a Monostate class through derivation"but that is obviously not the case for Borg, which, through multiple inheritance, makes such conversions trivial.Borg odds and ends The _ _getattr_ _ and _ _setattr_ _ special methods are not involved in Borg's operations. Therefore, you can define them independently in your subclass, for whatever other purposes you may require, or you may leave these special methods undefined. Either way is not a problem because Python does not call _ _setattr_ _ in the specific case of the rebinding of the instance's _ _dict_ _ attribute. Borg does not work well for classes that choose to keep some or all of their per-instance state somewhere other than in the instance's _ _dict_ _. So, in subclasses of Borg, avoid defining _ _slots_ _that's a memory-footprint optimization that would make no sense, anyway, since it's meant for classes that have a large number of instances, and Borg subclasses will effectively have just one instance! Moreover, instead of inheriting from built-in types such as list or dict, your Borg subclasses should use wrapping and automatic delegation, as shown previously Recipe 6.5. (I named this latter twist "DeleBorg," in my paper available at http://www.aleax.it/5ep.html.) Saying that Borg "is a Singleton" would be as silly as saying that a portico is an umbrella. Both serve similar purposes (letting you walk in the rain without getting wet)solve similar forces, in design pattern parlancebut since they do so in utterly different ways, they're not instances of the same pattern. If anything, as already mentioned, Borg has similarities to the Monostate alternative design pattern to Singleton. However, Monostate is a design pattern, while Borg is not; also, a Python Monostate could perfectly well exist without being a Borg. We can say that Borg is an idiom that makes it easy and effective to implement Monostate in Python. For reasons mysterious to me, people often conflate issues germane to Borg and Highlander with other, independent issues, such as access control and, particularly, access from multiple threads. If you need to control access to an object, that need is exactly the same whether there is one instance of that object's class or twenty of them, and whether or not those instances share state. A fruitful approach to problem-solving is known as divide and conquermaking problems easier to solve by splitting apart their different aspects. Making problems more difficult to solve by joining together several aspects must be an example of an approach known as unite and suffer! See Also Recipe 6.5; Recipe 6.15; Alex Martelli, "Five Easy Pieces: Simple Python Non-Patterns" (http://www.aleax.it/5ep.html). Recipe 6.17. Implementing the Null Object Design Pattern Credit: Dinu C. Gherman, Holger Krekel Problem You want to reduce the need for conditional statements in your code, particularly the need to keep checking for special cases. Solution The usual placeholder object for "there's nothing here" is None, but we may be able to do better than that by defining a class meant exactly to act as such a placeholder: class Null(object): """ Null objects always and reliably "do nothing." """ # optional optimization: ensure only one instance per subclass # (essentially just to save memory, no functional difference) def _ _new_ _(cls, *args, **kwargs): if '_inst' not in vars(cls): cls._inst = type._ _new_ _(cls, *args, **kwargs) return cls._inst def _ _init_ _(self, *args, **kwargs): pass def _ _call_ _(self, *args, **kwargs): return self def _ _repr_ _(self): return "Null( )" def _ _nonzero_ _(self): return False def _ _getattr_ _(self, name): return self def _ _setattr_ _(self, name, value): return self def _ _delattr_ _(self, name): return self Discussion You can use an instance of the Null class instead of the primitive value None. By using such an instance as a placeholder, instead of None, you can avoid many conditional statements in your code and can often express algorithms with little or no checking for special values. This recipe is a sample implementation of the Null Object Design Pattern. (See B. Woolf, "The Null Object Pattern" in Pattern Languages of Programming [PLoP 96, September 1996].) This recipe's Null class ignores all parameters passed when constructing or calling instances, as well as any attempt to set or delete attributes. Any call or attempt to access an attribute (or a method, since Python does not distinguish between the two, calling _ _getattr_ _ either way) returns the same Null instance (i.e., selfno reason to create a new instance). For example, if you have a computation such as: def compute(x, y): try: lots of computation here to return some appropriate object except SomeError: return Noneand you use it like this: for x in xs: for y in ys: obj = compute(x, y) if obj is not None: obj.somemethod(y, x) you can usefully change the computation to: def compute(x, y): try: lots of computation here to return some appropriate object except SomeError: return Null( ) and thus simplify its use down to: for x in xs: for y in ys: compute(x, y).somemethod(y, x) The point is that you don't need to check whether compute has returned a real result or an instance of Null: even in the latter case, you can safely and innocuously call on it whatever method you want. Here is another, more specific use case: log = err = Null( ) if verbose: log = open('/tmp/log', 'w') err = open('/tmp/err', 'w') log.write('blabla') err.write('blabla error') This obviously avoids the usual kind of "pollution" of your code from guards such as if verbose: strewn all over the place. You can now call log.write('bla'), instead of having to express each such call as if log is not None: log.write('bla'). In the new object model, Python does not call _ _getattr_ _ on an instance for any special methods needed to perform an operation on the instance (rather, it looks up such methods in the instance class' slots). You may have to take care and customize Null to your application's needs regarding operations on null objects, and therefore special methods of the null objects' class, either directly in the class' sources or by subclassing it appropriately. For example, with this recipe's Null, you cannot index Null instances, nor take their length, nor iterate on them. If this is a problem for your purposes, you can add all the special methods you need (in Null itself or in an appropriate subclass) and implement them appropriatelyfor example: class SeqNull(Null): def _ _len_ _(self): return 0 def _ _iter_ _(self): return iter(( )) def _ _getitem_ _(self, i): return self def _ _delitem_ _(self, i): return self def _ _setitem_ _(self, i, v): return self Similar considerations apply to several other operations.The key goal of Null objects is to provide an intelligent replacement for the often-used primitive value None in Python. (Other languages represent the lack of a value using either null or a null pointer.) These nobody-lives-here markers/placeholders are used for many purposes, including the important case in which one member of a group of otherwise similar elements is special. This usage usually results in conditional statements all over the place to distinguish between ordinary elements and the primitive null (e.g., None) value, but Null objects help you avoid that. Among the advantages of using Null objects are the following: Superfluous conditional statements can be avoided by providing a first-class object alternative for the primitive value None, thereby improving code readability. Null objects can act as placeholders for objects whose behavior is not yet implemented. Null objects can be used polymorphically with instances of just about any other class (perhaps needing suitable subclassing for special methods, as previously mentioned). Null objects are very predictable. The one serious disadvantage of Null is that it can hide bugs. If a function returns None, and the caller did not expect that return value, the caller most likely will soon thereafter try to call a method or perform an operation that None doesn't support, leading to a reasonably prompt exception and traceback. If the return value that the caller didn't expect is a Null, the problem might stay hidden for a longer time, and the exception and traceback, when they eventually happen, may therefore be harder to reconnect to the location of the defect in the code. Is this problem serious enough to make using Null inadvisable? The answer is a matter of opinion. If your code has halfway decent unit tests, this problem will not arise; while, if your code lacks decent unit tests, then using Null is the least of your problems. But, as I said, it boils down to a matter of opinions. I use Null very widely, and I'm extremely happy with the effect it has had on my productivity. The Null class as presented in this recipe uses a simple variant of the "Singleton" pattern (shown earlier in Recipe 6.15), strictly for optimization purposesnamely, to avoid the creation of numerous passive objects that do nothing but take up memory. Given all the previous remarks about customization by subclassing, it is, of course, crucial that the specific implementation of "Singleton" ensures a separate instance exists for each subclass of Null that gets instantiated. The number of subclasses will no doubt never be so high as to eat up substantial amounts of memory, and anyway this per-subclass distinction can be semantically crucial. See Also B. Woolf, "The Null Object Pattern" in Pattern Languages of Programming (PLoP 96, September 1996), http://www.cs.wustl.edu/~schmidt/PLoP-96/woolf1.ps.gz; Recipe 6.15. Recipe 6.18. Automatically Initializing Instance Variables from _ _init_ _ Arguments Credit: Peter Otten, Gary Robinson, Henry Crutcher, Paul Moore, Peter Schwalm, Holger Krekel Problem You want to avoid writing and maintaining _ _init_ _ methods that consist of almost nothing but a series of self.something = something assignments. Solution You can "factor out" the attribute-assignment task to an auxiliary function: def attributesFromDict(d): self = d.pop('self') for n, v in d.iteritems( ): setattr(self, n, v) Now, the typical boilerplate code for an _ _init_ _ method such as: def _ _init_ _(self, foo, bar, baz, boom=1, bang=2): self.foo = foo self.bar = bar self.baz = baz self.boom = boom self.bang = bang can become a short, crystal-clear one-liner: def _ _init_ _(self, foo, bar, baz, boom=1, bang=2): attributesFromDict(locals( )) Discussion As long as no additional logic is in the body of _ _init_ _, the dict returned by calling the built- in function locals contains only the arguments that were passed to _ _init_ _ (plus those arguments that were not passed but have default values). Function attributesFromDict extracts the object, relying on the convention that the object is always an argument named 'self', and then interprets all other items in the dictionary as names and values of attributes to set. A similar but simpler technique, not requiring an auxiliary function, is: def _ _init_ _(self, foo, bar, baz, boom=1, bang=2): self._ _dict_ _.update(locals( )) del self.self However, this latter technique has a serious defect when compared to the one presented in thisrecipe's Solution: by setting attributes directly into self._ _dict_ _ (through the latter's update method), it does not play well with properties and other advanced descriptors, while the approach in this recipe's Solution, using built-in setattr, is impeccable in this respect. attributesFromDict is not meant for use in an _ _init_ _ method that contains more code, and specifically one that uses some local variables, because attributesFromDict cannot easily distinguish, in the dictionary that is passed as its only argument d, between arguments of _ _init_ _ and other local variables of _ _init_ _. If you're willing to insert a little introspection in the auxiliary function, this limitation may be overcome: def attributesFromArguments(d): self = d.pop('self') codeObject = self._ _init_ _.im_func.func_code argumentNames = codeObject.co_varnames[1:codeObject.co_argcount] for n in argumentNames: setattr(self, n, d[n]) By extracting the code object of the _ _init_ _ method, function attributesFromArguments is able to limit itself to the names of _ _init_ _'s arguments. Your _ _init_ _ method can then call attributesFromArguments(locals( )), instead of attributesFromDict(locals( )), if and when it needs to continue, after the call, with more code that may define other local variables. The key limitation of attributesFromArguments is that it does not support _ _init_ _ having a last special argument of the **kw kind. Such support can be added, with yet more introspection, but it would require more black magic and complication than the functionality is probably worth. If you nevertheless want to explore this possibility, you can use the inspect module of the standard library, rather than the roll-your-own approach used in function attributeFromArguments, for introspection purposes. inspect.getargspec(self._ _init_ _) gives you both the argument names and the indication of whether self._ _init_ _ accepts a **kw form. See Recipe 6.19 for more information about function inspect.getargspec. Remember the golden rule of Python programming: "Let the standard library do it!" See Also Library Reference and Python in a Nutshell docs for the built-in function locals, methods of type dict, special method _ _init_ _, and introspection techniques (including module inspect). Recipe 6.19. Calling a Superclass _ _init_ _ Method If It Exists Credit: Alex Martelli Problem You want to ensure that _ _init_ _ is called for all superclasses that define it, and Python does not do this automatically. Solution As long as your class is new-style, the built-in super makes this task easy (if all superclasses' _ _init_ _ methods also use super similarly): class NewStyleOnly(A, B, C): def _ _init_ _(self): super(NewStyleOnly, self)._ _init_ _( ) initialization specific to subclass NewStyleOnly Discussion Classic classes are not recommended for new code development: they exist only to guarantee backwards compatibility with old versions of Python. Use new-style classes (deriving directly or indirectly from object) for all new code. The only thing you cannot do with a new-style class is to raise its instances as exception objects; exception classes must therefore be old style, but then, you do not need the functionality of this recipe for such classes. Since the rest of this recipe's Discussion is therefore both advanced and of limited applicability, you may want to skip it. Still, it may happen that you need to retrofit this functionality into a classic class, or, more likely, into a new-style class with some superclasses that do not follow the proper style of cooperative superclass method-calling with the built-in super. In such cases, you should first try to fix the problematic premisesmake all classes new style and make them use super properly. If you absolutely cannot fix things, the best you can do is to have your class loop over its base classesfor each base, check whether it has an _ _init_ _, and if so, then call it: class LookBeforeYouLeap(X, Y, Z): def _ _init_ _(self): for base in self_ _class_ _._ _bases_ _: if hasattr(base, '_ _init_ _'): base._ _init_ _(self) initialization specific to subclass LookBeforeYouLeap More generally, and not just for method _ _init_ _, we often want to call a method on an instance, or class, if and only if that method exists; if the method does not exist on that class or instance, we do nothing, or we default to another action. The technique shown in the "Solution", based on built-in super, is not applicable in general: it only works on superclasses of the current object, only if those superclasses also use super appropriately, and only if the method in question does exist in some superclass. Note that all new-style classes do have an _ _init_ _method: they all subclass object, and object defines _ _init_ _ (as a do-nothing function that accepts and ignores any arguments). Therefore, all new-style classes have an _ _init_ _ method, either by inheritance or by override. The LBYL technique shown in class LookBeforeYouLeap may be of help in more general cases, including ones that involve methods other than _ _init_ _. Indeed, LBYL may even be used together with super, for example, as in the following toy example: class Base1(object): def met(self): print 'met in Base1' class Der1(Base1): def met(self): s = super(Der1, self) if hasattr(s, 'met'): s.met( ) print 'met in Der1' class Base2(object): pass class Der2(Base2): def met(self): s = super(Der2, self) if hasattr(s, 'met'): s.met( ) print 'met in Der2' Der1( ).met( ) Der2( ).met( ) This snippet emits: met in Base1 met in Der1 met in Der2 The implementation of met has the same structure in both derived classes, Der1 (whose superclass Base1 does have a method named met) and Der2 (whose superclass Base1 doesn't have such a method). By binding a local name s to the result of super, and checking with hasattr that the superclass does have such a method before calling it, this LBYL structure lets you code in the same way in both cases. Of course, when coding a subclass, you do normally know which methods the superclasses have, and whether and how you need to call them. Still, this technique can provide a little extra flexibility for those occasions in which you need to slightly decouple the subclass from the superclass. The LBYL technique is far from perfect, though: a superclass might define an attribute named met, which is not callable or needs a different number of arguments. If your need for flexibility is so extreme that you must ward against such occurrences, you can extract the superclass' method object (if any) and check it with the getargspec function of standard library module inspect. While pushing this idea towards full generality can lead into rather deep complications, here is one example of how you might code a class with a method that calls the superclass' version of the same method only if the latter is callable without arguments: import inspect class Der(A, B, C, D): def met(self): s = super(Der, self) # get the superclass's bound-method object, or else None m = getattr(s, 'met', None) try: args, varargs, varkw, defaults = inspect.getargspec(m) except TypeError: # m is not a method, just ignore it pass else: # m is a method, do all its arguments have default values? if len(defaults) == len(args): # yes! so, call it: m( ) print 'met in Der' inspect.getargspec raises a TypeError if its argument is not a method or function, so we catch that case with a TRy/except statement, and if the exception occurs, we just ignore it with a do- nothing pass statement in the except clause. To simplify our code a bit, we do not first check separately with hasattr. Rather, we get the 'met' attribute of the superclass by calling getattr with a third argument of None. Thus, if the superclass does not have any attribute named 'met', m is set to None, later causing exactly the same TypeError that we have to catch (and ignore) anywaytwo birds with one stone. If the call to inspect.getargspec in the try clause does not raise a TypeError, execution continues with the else clause. If inspect.getargspec doesn't raise a TypeError, it returns a tuple of four items, and we bind each item to a local name. In this case, the ones we care about are args, a list of m's argument names, and defaults, a tuple of default values that m provides for its arguments. Clearly, we can call m without arguments if and only if m provides default values for all of its arguments. So, we check that there are just as many default values as arguments, by comparing the lengths of list args and tuple defaults, and call m only if the lengths are equal. No doubt you don't need such advanced introspection and such careful checking in most of the code you write, but, just in case you do, Python does supply all the tools you need to achieve it. See Also Docs for built-in functions super, getattr, and hasattr, and module inspect, in the Library Reference and Python in a Nutshell. Recipe 6.20. Using Cooperative Supercalls Concisely and Safely Credit: Paul McNett, Alex Martelli Problem You appreciate the cooperative style of multiple-inheritance coding supported by the super built- in, but you wish you could use that style in a more terse and concise way. Solution A good solution is a mixin classa class you can multiply inherit from, that uses introspection to allow more terse coding: import inspect class SuperMixin(object): def super(cls, *args, **kwargs): frame = inspect.currentframe(1) self = frame.f_locals['self'] methodName = frame.f_code.co_name method = getattr(super(cls, self), methodName, None) if inspect.ismethod(method): return method(*args, **kwargs) super = classmethod(super) Any class cls that inherits from class SuperMixin acquires a magic method named super: calling cls.super(args) from within a method named somename of class cls is a concise way to call super(cls, self).somename(args). Moreover, the call is safe even if no class that follows cls in Method Resolution Order (MRO) defines any method named somename. Discussion Here is a usage example: if _ _name_ _ == '_ _main_ _': class TestBase(list, SuperMixin): # note: no myMethod defined here pass class MyTest1(TestBase): def myMethod(self): print "in MyTest1" MyTest1.super( ) class MyTest2(TestBase): def myMethod(self): print "in MyTest2" MyTest2.super( ) class MyTest(MyTest1, MyTest2): def myMethod(self): print "in MyTest" MyTest.super( ) MyTest( ).myMethod( ) # emits: # in MyTest # in MyTest1 # in MyTest2 Python has been offering "new-style" classes for years, as a preferable alternative to the classic classes that you get by default. Classic classes exist only for backwards-compatibility with old versions of Python and are not recommended for new code. Among the advantages of new-style classes is the ease of calling superclass implementations of a method in a "cooperative" way that fully supports multiple inheritance, thanks to the super built-in. Suppose you have a method in a new-style class cls, which needs to perform a task and then delegate the rest of the work to the superclass implementation of the same method. The code idiom is: def somename(self, *args): ...some preliminary task... return super(cls, self).somename(*args) This idiom suffers from two minor issues: it's slightly verbose, and it also depends on a superclass offering a method somename. If you want to make cls less coupled to other classes, and therefore more robust, by removing the dependency, the code gets even more verbose: def somename(self, *args): ...some preliminary task... try: super_method = super(cls, self).somename except AttributeError: return None else: return super_method(*args) The mixin class SuperMixin shown in this recipe removes both issues. Just ensure cls inherits, directly or indirectly, from SuperMixin (alongside any other base classes you desire), and then you can code, concisely and robustly: def somename(self, *args): ...some preliminary task... return cls.super(*args) The classmethod SuperMixin.super relies on simple introspection to get the self object and the name of the method, then internally uses built-ins super and getattr to get the superclass method, and safely call it only if it exists. The introspection is performed through the handy inspect module of the standard Python library, making the whole task even simpler. See Also Library Reference and Python in a Nutshell docs on super, the new object model and MRO, the built-in getattr, and standard library module inspect; Recipe 20.12 for another recipe taking a very different approach to simplify the use of built-in super. Chapter 7. Persistence and Databases Introduction Recipe 7.1. Serializing Data Using the marshal Module Recipe 7.2. Serializing Data Using the pickle and cPickle Modules Recipe 7.3. Using Compression with Pickling Recipe 7.4. Using the cPickle Module on Classes and Instances Recipe 7.5. Holding Bound Methods in a Picklable Way Recipe 7.6. Pickling Code Objects Recipe 7.7. Mutating Objects with shelve Recipe 7.8. Using the Berkeley DB Database Recipe 7.9. Accesssing a MySQL Database Recipe 7.10. Storing a BLOB in a MySQL Database Recipe 7.11. Storing a BLOB in a PostgreSQL Database Recipe 7.12. Storing a BLOB in a SQLite Database Recipe 7.13. Generating a Dictionary Mapping Field Names to Column Numbers Recipe 7.14. Using dtuple for Flexible Accessto Query Results Recipe 7.15. Pretty-Printing the Contents of Database Cursors Recipe 7.16. Using a Single Parameter-Passing Style Across Various DB API Modules Recipe 7.17. Using Microsoft Jet via ADO Recipe 7.18. Accessing a JDBC Database from a Jython Servlet Recipe 7.19. Using ODBC to Get Excel Data with Jython Introduction Credit: Aaron Watters, Software Consultant There are three kinds of people in this world: those who can count and those who can't. However, there are only two kinds of computer programs: toy programs and programs that interact with some kind of persistent databases. That is to say, most real computer programs must retrieve stored information and record information for future use. These days, this description applies to almost every computer game, which can typically save and restore the state of the game at any time. So when I refer to toy programs, I mean programs written as exercises, or for the fun of programming. Nearly all real programs (such as programs that people get paid to write) have some persistent database storage/retrieval component. When I was a Fortran programmer in the 1980s, I noticed that although almost every program had to retrieve and store information, they almost always did it using home-grown methods. Furthermore, since the storage and retrieval parts of the program were the least interesting components from the programmer's point of view, these parts of the program were frequently implemented very sloppily and were hideous sources of intractable bugs. This repeated observation convinced me that the study and implementation of database systems sat at the core of programming pragmatics, and that the state of the art as I saw it then required much improvement. Later, in graduate school, I was delighted to find an impressive and sophisticated body of work relating to the implementation of database systems. The literature of database systems covered issues of concurrency, fault tolerance, distribution, query optimization, database design, and transaction semantics, among others. In typical academic fashion, many of the concepts had been elaborated to the point of absurdity (such as the silly notion of conditional multivalued dependencies), but much of the work was directly related to the practical implementation of reliable and efficient storage and retrieval systems. The starting point for much of this work was E.F. Codd's seminal paper, "A Relational Model of Data for Large Shared Data Banks."[1] [1] E.F. Codd, "A Relational Model of Data for Large Shared Data Banks," Communications of the ACM, 13, no. 6 (1970), pp. 377-87, http://www.acm.org/classics/nov95/toc.html. Among my fellow graduate students, and even among most of the faculty, the same body of knowledge was either disregarded or regarded with some scorn. Everyone recognized that knowledge of conventional relational technology could be lucrative, but they generally considered such knowledge to be on the same level as knowing how to write (or more importantly, maintain) COBOL programs. This situation was not helped by the fact that the emerging database interface standard, SQL (which is now very well established), looked like an extension of COBOL and bore little obvious relationship to any modern programming language. More than a decade later, there is little indication that anything will soon overtake SQL-based relational technology for the majority of data-based applications. In fact, relational-database technology seems more pervasive than ever. The largest software vendorsIBM, Microsoft, and Oracleall provide various relational-database implementations as crucial components of their core offerings. Other large software firms, such as SAP and PeopleSoft, essentially provide layers of software built on top of a relational-database core. Generally, relational databases have been augmented rather than replaced. Enterprise software- engineering dogma frequently espouses three-tier systems, in which the bottom tier is a carefully designed relational database, the middle tier defines a view of the database as business objects, and the top tier consists of applications or transactions that manipulate the business objects, with effects that ultimately translate to changes in the underlying relational tables. Microsoft's Open Database Connectivity (ODBC) standard provides a common programming APIfor SQL-based relational databases that permits programs to interact with many different database engines with no or few changes. For example, a Python program could be first implemented using Microsoft Jet[2] as a backend database for testing and debugging purposes. Once the program is stable, it can be put into production use, remotely accessing, say, a backend DB2 database on an IBM mainframe residing on another continent, by changing (at most) one line of code. [2] Microsoft Jet is commonly but erroneously known as the "Microsoft Access database." Access is a product that Microsoft sells for designing and implementing database frontends; Jet is a backend that you may download for free from Microsoft's web site. Relational databases are not appropriate for all applications. In particular, a computer game or engineering design tool that must save and restore sessions should probably use a more direct method of persisting the logical objects of the program than the flat tabular representation encouraged in relational-database design. However, even in domains such as engineering or scientific information, a hybrid approach that uses some relational methods is often advisable. For example, I have seen a complex relational-database schema for archiving genetic-sequencing informationin which the sequences show up as binary large objects (BLOBs)but a tremendous amount of important ancillary information can fit nicely into relational tables. But as the reader has probably surmised, I fear, I speak as a relational zealot. Within the Python world there are many ways of providing persistence and database functionality. My personal favorite is Gadfly, http://gadfly.sourceforge.net/, a simple and minimal SQL implementation that works primarily with in-memory databases. It is my favorite for no other reason than because it is mine, and its biggest advantage is that, if it becomes unworkable for you, it is easy to switch over to another, industrial-strength SQL engine. Many Gadfly users have started an application with Gadfly (because it was easy to use) and switched later (because they needed more). However, many people may prefer to start by using other SQL implementations such as MySQL, Microsoft Access, Oracle, Sybase, Microsoft SQL Server, SQLite, or others that provide the advantages of an ODBC interface (which Gadfly does not do). Python provides a standard interface for accessing relational databases: the Python DB Application Programming Interface (Py-DBAPI), originally designed by Greg Stein. Each underlying database API requires a wrapper implementation of the Py-DBAPI, and implementations are available for just about all underlying database interfaces, notably Oracle and ODBC. When the relational approach is overkill, Python provides built-in facilities for storing and retrieving data. At the most basic level, the programmer can manipulate files directly, as covered in Chapter 2. A step up from files, the marshal module allows programs to serialize data structures constructed from simple Python types (not including, e.g., classes or class instances). marshal has the advantage of being able to retrieve large data structures with blinding speed. The pickle and cPickle modules allow general storage of objects, including classes, class instances, and circular structures. cPickle is so named because it is implemented in C and is consequently quite fast, but it remains slower than marshal. For access to structured data in a somewhat human-readable form, it is also worth considering storing and retrieving data in XML format (taking advantage of Python's several XML parsing and generation modules), covered in Chapter 12but this option works best for write once, read many-type applications. Serialized data or XML representations may be stored in SQL databases to create a hybrid approach as well. While marshal and pickle provide basic serialization and deserialization of structures, the application programmer will frequently desire more functionality, such as transaction support and concurrency control. When the relational model doesn't fit the application, a direct object database implementation such as the Z-Object Database (ZODB) might be appropriatesee http://zope.org/Products/ZODB3.2. I must conclude with a plea to those who are dismissive of relational-database technology. Remember that it is successful for good reasons, and it might be worth considering. To paraphrase Churchill:text = """ Indeed, it has been said that democracy is the worst form of government, except for all those others that have been tried from time to time. """ import string for a, b in [("democracy", "SQL"), ("government", "database")]: text = string.replace(text, a, b) print text Recipe 7.1. Serializing Data Using the marshal Module Credit: Luther Blissett Problem You want to serialize and reconstruct a Python data structure whose items are fundamental Python objects (e.g., lists, tuples, numbers, and strings but no classes, instances, etc.) as fast as possible. Solution If you know that your data is composed entirely of fundamental Python objects (and you only need to support one version of Python, though possibly on several different platforms), the lowest-level, fastest approach to serializing your data (i.e., turning it into a string of bytes, and later reconstructing it from such a string) is via the marshal module. Suppose that data has only elementary Python data types as items, for example: data = {12:'twelve', 'feep':list('ciao'), 1.23:4+5j, (1,2,3):u'wer'} You can serialize data to a bytestring at top speed as follows: import marshal bytes = marshal.dumps(data) You can now sling bytes around as you wish (e.g., send it across a network, put it as a BLOB in a database, etc.), as long as you keep its arbitrary binary bytes intact. Then you can reconstruct the data structure from the bytestring at any time: redata = marshal.loads(bytes) When you specifically want to write the data to a disk file (as long as the latter is open for binarynot the default text modeinput/output), you can also use the dump function of the marshal module, which lets you dump several data structures to the same file one after the other: ouf = open('datafile.dat', 'wb') marshal.dump(data, ouf) marshal.dump('some string', ouf) marshal.dump(range(19), ouf) ouf.close( ) You can later recover from datafile.dat the same data structures you dumped into it, in the same sequence: inf = open('datafile.dat', 'rb') a = marshal.load(inf) b = marshal.load(inf) c = marshal.load(inf) inf.close( )Discussion Python offers several ways to serialize data (meaning to turn the data into a string of bytes that you can save on disk, put in a database, send across the network, etc.) and corresponding ways to reconstruct the data from such serialized forms. The lowest-level approach is to use the marshal module, which Python uses to write its bytecode files. marshal supports only elementary data types (e.g., dictionaries, lists, tuples, numbers, and strings) and combinations thereof. marshal does not guarantee compatibility from one Python release to another, so data serialized with marshal may not be readable if you upgrade your Python release. However, marshal does guarantee independence from a specific machine's architecture, so it is guaranteed to work if you're sending serialized data between different machines, as long as they are all running the same version of Pythonsimilar to how you can share compiled Python bytecode files in such a distributed setting. marshal's dumps function accepts any suitable Python data structure and returns a bytestring representing it. You can pass that bytestring to the loads function, which will return another Python data structure that compares equal (==) to the one you originally dumped. In particular, the order of keys in dictionaries is arbitrary in both the original and reconstructed data structures, but order in any kind of sequence is meaningful and is thus preserved. In between the dumps and loads calls, you can subject the bytestring to any procedure you wish, such as sending it over the network, storing it into a database and retrieving it, or encrypting and decrypting it. As long as the string's binary structure is correctly restored, loads will work fine on it (as stated previously, this is guaranteed only if you use loads under the same Python release with which you originally executed dumps). When you specifically need to save the data to a file, you can also use marshal's dump function, which takes two arguments: the data structure you're dumping and the open file object. Note that the file must be opened for binary I/O (not the default, which is text I/O) and can't be a file- like object, as marshal is quite picky about it being a true file. The advantage of dump is that you can perform several calls to dump with various data structures and the same open file object: each data structure is then dumped together with information about how long the dumped bytestring is. As a consequence, when you later open the file for binary reading and then call marshal.load, passing the file as the argument, you can reload each previously dumped data structure sequentially, one after the other, at each call to load. The return value of load, like that of loads, is a new data structure that compares equal to the one you originally dumped. (Again, dump and load work within one Python releaseno guarantee across releases.) Those accustomed to other languages and libraries offering "serialization" facilities may be wondering if marshal imposes substantial practical limits on the size of objects you can serialize or deserialize. Answer: Nope. Your machine's memory might, but as long as everything fits comfortably in memory, marshal imposes practically no further limit. See Also Recipe 7.2 for cPickle, the big brother of marshal; documentation on the marshal standard library module in the Library Reference and in Python in a Nutshell. Recipe 7.2. Serializing Data Using the pickle and cPickle Modules Credit: Luther Blissett Problem You want to serialize and reconstruct, at a reasonable speed, a Python data structure, which may include both fundamental Python object as well as classes and instances. Solution If you don't want to assume that your data is composed only of fundamental Python objects, or you need portability across versions of Python, or you need to transmit the serialized form as text, the best way of serializing your data is with the cPickle module. (The pickle module is a pure-Python equivalent and totally interchangeable, but it's slower and not worth using except if you're missing cPickle.) For example, say you have: data = {12:'twelve', 'feep':list('ciao'), 1.23:4+5j, (1,2,3):u'wer'} You can serialize data to a text string: import cPickle text = cPickle.dumps(data) or to a binary string, a choice that is faster and takes up less space: bytes = cPickle.dumps(data, 2) You can now sling text or bytes around as you wish (e.g., send across a network, include as a BLOB in a databasesee Recipe 7.10, Recipe 7.11, and Recipe 7.12) as long as you keep text or bytes intact. In the case of bytes, it means keeping the arbitrary binary bytes intact. In the case of text, it means keeping its textual structure intact, including newline characters. Then you can reconstruct the data at any time, regardless of machine architecture or Python release: redata1 = cPickle.loads(text) redata2 = cPickle.loads(bytes) Either call reconstructs a data structure that compares equal to data. In particular, the order of keys in dictionaries is arbitrary in both the original and reconstructed data structures, but order in any kind of sequence is meaningful, and thus it is preserved. You don't need to tell cPickle.loads whether the original dumps used text mode (the default, also readable by some very old versions of Python) or binary (faster and more compact)loads figures it out by examining its argument's contents. When you specifically want to write the data to a file, you can also use the dump function of the cPickle module, which lets you dump several data structures to the same file one after the other:ouf = open('datafile.txt', 'w') cPickle.dump(data, ouf) cPickle.dump('some string', ouf) cPickle.dump(range(19), ouf) ouf.close( ) Once you have done this, you can recover from datafile.txt the same data structures you dumped into it, one after the other, in the same order: inf = open('datafile.txt') a = cPickle.load(inf) b = cPickle.load(inf) c = cPickle.load(inf) inf.close( ) You can also pass cPickle.dump a third argument with a value of 2 to tell cPickle.dump to serialize the data in binary form (faster and more compact), but the data file must then be opened for binary I/O, not in the default text mode, both when you originally dump to the file and when you later load from the file. Discussion Python offers several ways to serialize data (i.e., make the data into a string of bytes that you can save on disk, save in a database, send across the network, etc.) and corresponding ways to reconstruct the data from such serialized forms. Typically, the best approach is to use the cPickle module. A pure-Python equivalent, called pickle (the cPickle module is coded in C as a Python extension) is substantially slower, and the only reason to use it is if you don't have cPickle (e.g., with a Python port onto a mobile phone with tiny storage space, where you saved every byte you possibly could by installing only an indispensable subset of Python's large standard library). However, in cases where you do need to use pickle, rest assured that it is completely interchangeable with cPickle: you can pickle with either module and unpickle with the other one, without any problems whatsoever. cPickle supports most elementary data types (e.g., dictionaries, lists, tuples, numbers, strings) and combinations thereof, as well as classes and instances. Pickling classes and instances saves only the data involved, not the code. (Code objects are not even among the types that cPickle knows how to serialize, basically because there would be no way to guarantee their portability across disparate versions of Python. See Recipe 7.6 for a way to serialize code objects, as long as you don't need the cross-version guarantee.) See Recipe 7.4 for more about pickling classes and instances. cPickle guarantees compatibility from one Python release to another, as well as independence from a specific machine's architecture. Data serialized with cPickle will still be readable if you upgrade your Python release, and pickling is also guaranteed to work if you're sending serialized data between different machines. The dumps function of cPickle accepts any Python data structure and returns a text string representing it. If you call dumps with a second argument of 2, dumps returns an arbitrary bytestring instead: the operation is faster, and the resulting string takes up less space. You can pass either the text or the bytestring to the loads function, which will return another Python data structure that compares equal (==) to the one you originally dumped. In between the dumps and loads calls, you can subject the text or bytestring to any procedure you wish, such as sending it over the network, storing it in a database and retrieving it, or encrypting and decrypting it. As long as the string's textual or binary structure is correctly restored, loads will work fine on it (even across platforms and in future releases). If you need to produce data readable by old (pre- 2.3) versions of Python, consider using 1 as the second argument: operation will be slower, andthe resulting strings will not be as compact as those obtained by using 2, but the strings will be unpicklable by old Python versions as well as current and future ones. When you specifically need to save the data into a file, you can also use cPickle's dump function, which takes two arguments: the data structure you're dumping and the open file or file-like object. If the file is opened for binary I/O, rather than the default (text I/O), then by giving dump a third argument of 2, you can ask for binary format, which is faster and takes up less space (again, you can also use 1 in this position to get a binary format that's neither as compact nor as fast, but is understood by old, pre-2.3 Python versions too). The advantage of dump over dumps is that, with dump, you can perform several calls, one after the other, with various data structures and the same open file object. Each data structure is then dumped with information about how long the dumped string is. Consequently, when you later open the file for reading (binary reading, if you asked for binary format) and then repeatedly call cPickle.load, passing the file as the argument, each data structure previously dumped is reloaded sequentially, one after the other. The return value of load, like that of loads, is a new data structure that compares equal to the one you originally dumped. Those accustomed to other languages and libraries offering "serialization" facilities may be wondering whether pickle imposes substantial practical limits on the size of objects you can serialize or deserialize. Answer: Nope. Your machine's memory might, but as long as everything fits comfortably in memory, pickle practically imposes no further limit. See Also Recipe 7.2 and Recipe 7.4; documentation for the standard library module cPickle in the Library Reference and Python in a Nutshell. Recipe 7.3. Using Compression with Pickling Credit: Bill McNeill, Andrew Dalke Problem You want to pickle generic Python objects to and from disk in a compressed form. Solution Standard library modules cPickle and gzip offer the needed functionality; you just need to glue them together appropriately: import cPickle, gzip def save(filename, *objects): ''' save objects into a compressed diskfile ''' fil = gzip.open(filename, 'wb') for obj in objects: cPickle.dump(obj, fil, proto=2) fil.close( ) def load(filename): ''' reload objects from a compressed diskfile ''' fil = gzip.open(filename, 'rb') while True: try: yield cPickle.load(fil) except EOFError: break fil.close( ) Discussion Persistence and compression, as a general rule, go well together. cPickle protocol 2 saves Python objects quite compactly, but the resulting files can still compress quite well. For example, on my Linux box, open('/usr/dict/share/words').readlines( ) produces a list of over 45,000 strings. Pickling that list with the default protocol 0 makes a disk file of 972 KB, while protocol 2 takes only 716 KB. However, using both gzip and protocol 2, as shown in this recipe, requires only 268 KB, saving a significant amount of space. As it happens, protocol 0 produces a more compressible file in this case, so that using gzip and protocol 0 would save even more space, taking only 252 KB on disk. However, the difference between 268 and 252 isn't all that meaningful, and protocol 2 has other advantages, particularly when used on instances of new- style classes, so I recommend the mix I use in the functions shown in this recipe. Whatever protocol you choose to save your data, you don't need to worry about it when you're reloading the data. The protocol is recorded in the file together with the data, so cPickle.load can figure out by itself all it needs. Just pass it an instance of a file or pseudo-file object with a read method, and cPickle.load returns each object that was pickled to the file, one after the other, and raises EOFError when the file's done. In this recipe, we wrap a generator around cPickle.load, so you can simply loop over all recovered objects with a for statement, or, depending on what you need, you can use some call such as list(load('somefile.gz')) to get a list with all recovered objects as its items.See Also Modules gzip and cPickle in the Library Reference. Recipe 7.4. Using the cPickle Module on Classes and Instances Credit: Luther Blissett Problem You want to save and restore class and instance objects using the cPickle module. Solution You often need no special precautions to use cPickle on your classes and their instances. For example, the following works fine: import cPickle class ForExample(object): def _ _init_ _(self, *stuff): self.stuff = stuff anInstance = ForExample('one', 2, 3) saved = cPickle.dumps(anInstance) reloaded = cPickle.loads(saved) assert anInstance.stuff == reloaded.stuff However, sometimes there are problems: anotherInstance = ForExample(1, 2, open('three', 'w')) wontWork = cPickle.dumps(anotherInstance) This snippet causes a TypeError: "can't pickle file objects" exception, because the state of anotherInstance includes a file object, and file objects cannot be pickled. You would get exactly the same exception if you tried to pickle any other container that includes a file object among its items. However, in some cases, you may be able to do something about it: class PrettyClever(object): def _ _init_ _(self, *stuff): self.stuff = stuff def _ _getstate_ _(self): def normalize(x): if isinstance(x, file): return 1, (x.name, x.mode, x.tell( )) return 0, x return [ normalize(x) for x in self.stuff ] def _ _setstate_ _(self, stuff): def reconstruct(x): if x[0] == 0: return x[1] name, mode, offs = x[1] openfile = open(name, mode) openfile.seek(offs) return openfile self.stuff = tuple([reconstruct(x) for x in stuff]) By defining the _ _getstate_ _ and _ _setstate_ _ special methods in your class, you gain fine-grained control about what, exactly, your class' instances consider to be their state. As long as you can define such state in picklable terms, and reconstruct your instances from the unpickled state in some way that is sufficient for your application, you can make your instances themselves picklable and unpicklable in this way. Discussion cPickle dumps class and function objects by name (i.e., through their module's name and their name within the module). Thus, you can dump only classes defined at module level (not inside other classes and functions). Reloading such objects requires the respective modules to be available for import. Instances can be saved and reloaded only if they belong to such classes. In addition, the instance's state must also be picklable. By default, an instance's state is the contents of the instance's _ _dict_ _, plus whatever state the instance may get from the built-in type the instance's class inherits from, if any. For example, an instance of a new-style class that subclasses list includes the list items as part of the instance's state. cPickle also handles instances of new-style classes that define or inherit a class attribute named _ _slots_ _ (and therefore hold some or all per-instance state in those predefined slots, rather than in a per-instance _ _dict_ _). Overall, cPickle's default approach is often quite sufficient and satisfactory. Sometimes, however, you may have nonpicklable attributes or items as part of your instance's state (as cPickle defines such state by default, as explained in the previous paragraph). In this recipe, for example, I show a class whose instances hold arbitrary stuff, which may include open file objects. To handle this case, your class can define the special method _ _getstate_ _. cPickle calls that method on your object, if your object's class defines it or inherits it, instead of going directly for the object's _ _dict_ _ (or possibly _ _slots_ _ and/or built-in type bases). Normally, when you define the _ _getstate_ _ method, you define the _ _setstate_ _ method as well, as shown in this recipe's Solution. _ _getstate_ _ can return any picklable object, and that object gets pickled, and later, at unpickling time, passed as _ _setstate_ _'s argument. In this recipe's Solution, _ _getstate_ _ returns a list that's similar to the instance's default state (attribute self.stuff), except that each item is turned into a tuple of two items. The first item in the pair can be set to 0 to indicate that the second one will be taken verbatim, or 1 to indicate that the second item will be used to reconstruct an open file. (Of course, the reconstruction may fail or be unsatisfactory in several ways. There is no general way to save an open file's state, which is why cPickle itself doesn't even try. But in the context of our application, we can assume that the given approach will work.) When reloading the instance from pickled form, cPickle calls _ _setstate_ _ with the list of pairs, and _ _setstate_ _ can reconstruct self.stuff by processing each pair appropriately in its nested reconstruct function. This scheme can clearly generalize to getting and restoring state that may contain various kinds of normally unpicklable objectsjust be sure to use different numbers to tag each of the various kinds of "nonverbatim" pairs you need to support. In one particular case, you can define _ _getstate_ _ without defining _ _setstate_ _: _ _getstate_ _ must then return a dictionary, and reloading the instance from pickled form uses that dictionary just as the instance's _ _dict_ _ would normally be used. Not running your own code at reloading time is a serious hindrance, but it may come in handy when you want to use _ _getstate_ _, not to save otherwise unpicklable state but rather as an optimization. Typically, this optimization opportunity occurs when your instance caches results that it can recompute if they're absent, and you decide it's best not to store the cache as a part of the instance's state. In this case, you should define _ _getstate_ _ to return a dictionary that's the indispensable subset of the instance's _ _dict_ _. (See Recipe 4.13) for a simple and handy way to "subset adictionary".) Defining _ _getstate_ _ (and then, normally, also _ _setstate_ _) also gives you a further important bonus, besides the pickling support: if a class offers these methods but doesn't offer special methods _ _copy_ _ or _ _deepcopy_ _, then the methods are also used for copying, both shallowly and deeply, as well as for serializing. The state data returned by _ _getstate_ _ is deep-copied if and only if the object is being dee-copied, but, other than this distinction, shallow and deep copies work very similarly when they are implemented through _ _getstate_ _. See Recipe 4.1 for more information about how a class can control the way its instances are copied, shallowly or deeply. With either the default pickling/unpickling approach, or your own _ _getstate_ _ and _ _setstate_ _, the instance's special method _ _init_ _ is not called when the instance is getting unpickled. If the most convenient way for you to reconstruct an instance is to call the _ _init_ _ method with appropriate parameters, then you may want to define the special method _ _getinitargs_ _, instead of _ _getstate_ _. In this case, cPickle calls this method without arguments: the method must return a pickable tuple, and at unpickling time, cPickle calls _ _init_ _ with the arguments that are that tuple's items. _ _getinitargs_ _, like _ _getstate_ _ and _ _setstate_ _, can also be used for copying. The Library Reference for the pickle and copy_reg modules details even subtler things you can do when pickling and unpickling, as well as the thorny security issues that are likely to arise if you ever stoop to unpickling data from untrusted sources. (Executive summary: don't do thatthere is no way Python can protect you if you do.) However, the techniques I've discussed here should suffice in almost all practical cases, as long as the security aspects of unpickling are not a problem (and if they are, the only practical suggestion is: forget pickling!). See Also Recipe 7.2; documentation for the standard library module cPickle in the Library Reference and Python in a Nutshell. Recipe 7.5. Holding Bound Methods in a Picklable Way Credit: Peter Cogolo Problem You need to pickle an object, but that object holds (as an attribute or item) a bound method of another object, and bound methods are not picklable. Solution Say you have the following objects: import cPickle class Greeter(object): def _ _init_ _(self, name): self.name = name def greet(self): print 'hello', self.name class Repeater(object): def _ _init_ _(self, greeter): self.greeter = greeter def greet(self): self.greeter( ) self.greeter( ) r = Repeater(Greeter('world').greet) Were it not for the fact that r holds a bound method as its greeter attribute, you could pickle r very simply: s = cPickle.dumps(r) However, upon encountering the bound method, this call to cPickle.dumps raises a TypeError. One simple solution is to have each instance of class Repeater hold, not a bound method directly, but rather a picklable wrapper to it. For example: class picklable_boundmethod(object): def _ _init_ _(self, mt): self.mt = mt def _ _getstate_ _(self): return self.mt.im_self, self.mt.im_func._ _name_ _ def _ _setstate_ _(self, (s,fn)): self.mt = getattr(s, fn) def _ _call_ _(self, *a, **kw): return self.mt(*a, **kw) Now, changing Repeater._ _init_ _'s body to self.greeter = picklable_boundmethod(greeter) makes the previous snippet work.Discussion The Python Standard Library pickle module (just like its faster equivalent cousin cPickle) pickles functions and classes by namethis implies, in particular, that only functions defined at the top level of a module can be pickled (the pickling of such a function, in practice, contains just the names of the module and function). If you have a graph of objects that hold each other, not directly, but via one another's bound methods (which is often a good idea in Python), this limitation can make the whole graph unpicklable. One solution might be to teach pickle how to serialize bound methods, along the same lines as described in Recipe 7.6. Another possible solution is to define appropriate _ _getstate_ _ and _ _setstate_ _ methods to turn bound methods into something picklable at dump time and rebuild them at load time, along the lines described in Recipe 7.4. However, this latter possibility is not a good factorization when you have several classes whose instances hold bound methods. This recipe pursues a simpler idea, based on holding bound methods, not directly, but via the picklable_boundmethod wrapper class. picklable_boundmethod is written under the assumption that the only thing you usually do with a bound method is to call it, so it only delegates _ _call_ _ functionality specifically. (You could, in addition, also use _ _getattr_ _, in order to delegate other attribute accesses.) In normal operation, the fact that you're holding an instance of picklable_boundmethod rather than holding the bound method object directly is essentially transparent. When pickling time comes, special method _ _getstate_ _ of picklable_boundmethod comes into play, as previously covered in Recipe 7.4. In the case of picklable_boundmethod, _ _getstate_ _ returns the object to which the bound method belongs and the function name of the bound method. Later, at unpickling time, _ _setstate_ _ recovers an equivalent bound method from the reconstructed object by using the getattr built-in for that name. This approach isn't infallible because an object might hold its methods under assumed names (different from the real function names of the methods). However, assuming you're not specifically doing something weird for the specific purpose of breaking picklable_boundmethod's functionality, you shouldn't ever run into this kind of obscure problem! See Also Library Reference and Python in a Nutshell docs for modules pickle and cPickle, bound-method objects, and the getattr built-in. Recipe 7.6. Pickling Code Objects Credit: Andres Tremols, Peter Cogolo Problem You want to be able to pickle code objects, but this functionality is not supported by the standard library's pickling modules. Solution You can extend the abilities of the pickle (or cPickle) module by using module copy_reg. Just make sure the following module has been imported before you pickle code objects, and has been imported, or is available to be imported, when you're unpickling them: import new, types, copy_reg def code_ctor(*args): # delegate to new.code the construction of a new code object return new.code(*args) def reduce_code(co): # a reductor function must return a tuple with two items: first, the # constructor function to be called to rebuild the argument object # at a future de-serialization time; then, the tuple of arguments # that will need to be passed to the constructor function. if co.co_freevars or co.co_cellvars: raise ValueError, "Sorry, cannot pickle code objects from closures" return code_ctor, (co.co_argcount, co.co_nlocals, co.co_stacksize, co.co_flags, co.co_code, co.co_consts, co.co_names, co.co_varnames, co.co_filename, co.co_name, co.co_firstlineno, co.co_lnotab) # register the reductor to be used for pickling objects of type 'CodeType' copy_reg.pickle(types.CodeType, reduce_code) if _ _name_ _ == '_ _main_ _': # example usage of our new ability to pickle code objects import cPickle # a function (which, inside, has a code object, of course) def f(x): print 'Hello,', x # serialize the function's code object to a string of bytes pickled_code = cPickle.dumps(f.func_code) # recover an equal code object from the string of bytes recovered_code = cPickle.loads(pickled_code) # build a new function around the rebuilt code object g = new.function(recovered_code, globals( )) # check what happens when the new function gets called g('world') Discussion The Python Standard Library pickle module (just like its faster equivalent cousin cPickle) pickles functions and classes by name. There is no pickling of the code objects containing thecompiled bytecode that, when run, determines almost every aspect of functions' (and methods') behavior. In some situations, you'd rather pickle everything by value, so that all the relevant stuff can later be retrieved from the pickle, rather than having to have module files around for some of it. Sometimes you can solve such problems by using marshaling rather than pickling, since marshal does let you serialize code objects, but marshal has limitations on many other issues. For example, you cannot marshal instances of classes you have coded. (Once you're serializing code objects, which are specific to a given version of Python, pickle will share one key limitation of marshal: no guaranteed ability to save and later reload data across different versions of Python.) An alternative approach is to take advantage of the possibility, which the Python Standard Library allows, to extend the set of types known to pickle. Basically, you can "teach" pickle how to save and reload code objects; this, in turn, lets you pickle by value, rather than "by name", such objects as functions and classes. (The code in this recipe's Solution under the if _ _name_ _ == '_ _main_ _' guard essentially shows how to extend pickle for a function.) To teach pickle about some new type, use module copy_reg, which is also part of the Python Standard Library. Through function copy_reg.pickle, you register the reduction function to use for instances of a given type. A reduction function takes as its argument an instance to be pickled and returns a tuple with two items: a constructor function, which will be called to reconstruct the instance, and a tuple of arguments, which will be passed to the constructor function. (A reduction function may also return other kinds of results, but for this recipe's purposes a two-item tuple suffices.) The module in this recipe defines function reduce_code, then registers it as the reduction function for objects of type types.CodeTypethat is, code objects. When reduce_code gets called, it first checks whether its code object co comes from a closure (functions nested inside each other), because it just can't deal with this eventualityI've been unable to find a way that works, so in this case, reduce_code just raises an exception to let the user know about the problem. In normal cases, reduce_code returns code_ctor as the constructor and a tuple made up of all of co's attributes as the arguments tuple for the constructor. When a code object is reloaded from a pickle, code_ctor gets called with those arguments and simply passes the call on to the new.code callable, which is the true constructor for code arguments. Unfortunately, reduce_code cannot return new.code itself as the first item in its result tuple, because new.code is a built-in (a C- coded callable) but is not available through a built-in name. So, basically, the role of code_ctor is to provide a name for the (by-name) pickling of new.code. The if _ _name_ _ == '_ _main_ _' part of the recipe provides a typical toy usage exampleit pickles a code object to a string, recovers a copy of it from the pickle string, and builds and calls a function around that code object. A more typical use case for this recipe's functionality, of course, will do the pickling in one script and the unpickling in another. Assume that the module in this recipe has been saved as file reco.py somewhere on Python's sys.path, so that it can be imported by Python scripts and other modules. You could then have a script that imports reco and thus becomes able to pickle code objects, such as: import reco, pickle def f(x): print 'Hello,', x pickle.dump(f.func_code, open('saved.pickle','wb')) To unpickle and use that code object, an example script might be: import new, cPickle c = cPickle.load(open('saved.pickle','rb')) g = new.function(c, globals( )) g('world') Note that the second script does not need to import recothe import will happen automaticallywhen needed (part of the information that pickle saves in saved.pickle is that, in order to reconstruct the pickled object therein, it needs to call reco.code_ctor; so, it also knows it needs to import reco). I'm also showing that you can use modules pickle and cPickle interchangeably. Pickle is faster, but there are no other differences, and in particular, you can use one module to pickle objects and the other one to unpickle them, if you wish. See Also Modules pickle, cPickle, and copy_reg in the Library Reference and Python in a Nutshell. Recipe 7.7. Mutating Objects with shelve Credit: Luther Blissett Problem You are using the standard module shelve. Some of the values you have shelved are mutable objects, and you need to mutate these objects. Solution The shelve module offers a kind of persistent dictionaryan important niche between the power of relational-database engines and the simplicity of marshal, pickle, dbm, and similar file formats. However, you should be aware of a typical trap you need to avoid when using shelve. Consider the following interactive Python session: >>> import shelve >>> # Build a simple sample shelf >>> she = shelve.open('try.she', 'c') >>> for c in 'spam': she[c] = {c:23} ... >>> for c in she.keys( ): print c, she[c] ... p {'p': 23} s {'s': 23} a {'a': 23} m {'m': 23} >>> she.close( ) We've created the shelve file, added some data to it, and closed it. Goodnow we can reopen it and work with it: >>> she=shelve.open('try.she', 'c') >>> she['p'] {'p': 23} >>> she['p']['p'] = 42 >>> she['p'] {'p': 23} What's going on here? We just set the value to 42, but our setting didn't take in the shelve object! The problem is that we were working with a temporary object that shelve gave us, not with the "real thing". shelve, when we open it with default options, like here, doesn't track changes to such temporary objects. One reasonable solution is to bind a name to this temporary object, do our mutation, and then assign the mutated object back to the appropriate item of shelve: >>> a = she['p'] >>> a['p'] = 42 >>> she['p'] = a >>> she['p']{'p': 42} >>> she.close( ) We can verify that the change was properly persisted: >>> she=shelve.open('try.she','c') >>> for c in she.keys( ): print c,she[c] ... p {'p': 42} s {'s': 23} a {'a': 23} m {'m': 23} A simpler solution is to open the shelve object with the writeback option set to TRue: >>> she = shelve.open('try.she', 'c', writeback=True) The writeback option instructs shelve to keep track of all the objects it gets from the file and write them all back to the file before closing it, just in case they have been modified in the meantime. While simple, this approach can be quite expensive, particularly in terms of memory consumption. Specifically, if we read many objects from a shelve object opened with writeback=True, even if we only modify a few of them, shelve is going to keep them all in memory, since it can't tell in advance which one we may be about to modify. The previous approach, where we explicitly take responsibility to notify shelve of any changes (by assigning the changed objects back to the place they came from), requires more care on our part, but repays that care by giving us much better performance. Discussion The standard Python module shelve can be quite convenient in many cases, but it hides a potentially nasty trap, admittedly well documented in Python's online docs but still easy to miss. Suppose you're shelving mutable objects, such as dictionaries or lists. Naturally, you are quite likely to want to mutate some of those objectsfor example, by calling mutating methods (append on a list, update on a dictionary, etc.) or by assigning a new value to an item or attribute of the object. However, when you do this, the change doesn't occur in the shelve object. This is because we actually mutate a temporary object that the shelve object has given us as the result of shelve's own _ _getitem_ _ method, but the shelve object, by default, does not keep track of that temporary object, nor does it care about it once it returns it to us. As shown in the recipe, one solution is to bind a name to the temporary object obtained by keying into the shelf, doing whatever mutations are needed to the object via the name, then assigning the newly mutated object back to the appropriate item of the shelve object. When you assign to a shelve object's item, the shelve object's _ _setitem_ _ method gets invoked, and it appropriately updates the shelve object itself, so that the change does occur. Alternatively, you can add the flag writeback=True at the time you open the shelve object, and then shelve keeps track of every object it hands you, saving them all back to disk at the end. This approach may save you quite a bit of fussy and laborious coding, but take care: if you read many items of the shelve object and only modify a few of them, the writeback approach can be exceedingly costly, particularly in terms of memory consumption. When opened with writeback=True, shelve will keep in memory any item it has ever handed you, and save them all to disk at the end, since it doesn't have a reliable way to tell which items you may be about to modify, nor, in general, even which items you have actually modified by the time you close the shelve object. The recommended approach, unless you're going to modify just about every item you read (or unless the shelve object in question is small enough compared with your available memory that you don't really care), is the previous one: bind a name to the items that you getfrom a shelve object with intent to modify them, and assign each item back into the shelve object once you're done mutating that item. See Also Recipe 7.1 and Recipe 7.2 for alternative serialization approaches; documentation for the shelve standard library module in the Library Reference and Python in a Nutshell. Recipe 7.8. Using the Berkeley DB Database Credit: Farhad Fouladi Problem You want to persist some data, exploiting the simplicity and good performance of the Berkeley DB database library. Solution If you have previously installed Berkeley DB on your machine, the Python Standard Library comes with package bsddb (and optionally bsddb3, to access Berkeley DB release 3.2 databases) to interface your Python code with Berkeley DB. To get either bsddb or, lacking it, bsddb3, use a try/except on import: try: from bsddb import db # first try release 4 except ImportError: from bsddb3 import db # not there, try release 3 instead print db.DB_VERSION_STRING # emits, e.g: Sleepycat Software: Berkeley DB 4.1.25: (December 19, 2002) To create a database, instantiate a db.DB object, then call its method open with appropriate parameters, such as: adb = db.DB( ) adb.open('db_filename', dbtype=db.DB_HASH, flags=db.DB_CREATE) db.DB_HASH is just one of several access methods you may choose when you create a databasea popular alternative is db.DB_BTREE, to use B+tree access (handy if you need to get records in sorted order). You may make an in-memory database, without an underlying file for persistence, by passing None instead of a filename as the first argument to the open method. Once you have an open instance of db.DB, you can add records, each composed of two strings, key and data: for i, w in enumerate('some words for example'.split( )): adb.put(w, str(i)) You can access records via a cursor on the database: def irecords(curs): record = curs.first( ) while record: yield record record = curs.next( ) for key, data in irecords(adb.cursor( )): print 'key=%r, data=%r' % (key, data) # emits (the order may vary):# key='some', data='0' # key='example', data='3' # key='words', data='1' # key='for', data='2' When you're done, you close the database: adb.close( ) At any future time, in the same or another Python program, you can reopen the database by giving just its filename as the argument to the open method of a newly created db.DB instance: the_same_db = db.DB( ) the_same_db.open('db_filename') and work on it again in the same ways: the_same_db.put('skidoo', '23') # add a record the_same_db.put('words', 'sweet') # replace a record for key, data in irecords(the_same_db.cursor( )): print 'key=%r, data=%r' % (key, data) # emits (the order may vary): # key='some', data='0' # key='example', data='3' # key='words', data='sweet' # key='for', data='2' # key='skidoo', data='23' Again, remember to close the database when you're done: the_same_db.close( ) Discussion The Berkeley DB is a popular open source database. It does not support SQL, but it's simple to use, offers excellent performance, and gives you a lot of control over exactly what happens, if you care to exert it, through a huge array of options, flags, and methods. Berkeley DB is just as accessible from many other languages as from Python: for example, you can perform some changes or queries with a Python program, and others with a separate C program, on the same database file, using the same underlying open source library that you can freely download from Sleepycat. The Python Standard Library shelve module can use the Berkeley DB as its underlying database engine, just as it uses cPickle for serialization. However, shelve does not let you take advantage of the ability to access a Berkeley DB database file from several different languages, exactly because the records are strings produced by pickle.dumps, and languages other than Python can't easily deal with them. Accessing the Berkeley DB directly with bsddb also gives you access to many advanced functionalities of the database engine that shelve simply doesn't support.A Database, or pickle . . . or Both? The use cases for pickle or marshal, and those for databases such as Berkeley DB or relational databases, are rather different, though they do overlap somewhat. pickle (and marshal even more so) is essentially about serialization: you turn Python objects into BLOBs that you may transmit or store, and later receive or retrieve. Data thus serialized is meant to be reloaded into Python objects, basically only by Python applications. pickle has nothing to say about searching or selecting specific objects or parts of them. Databases (Berkeley DB, relational DBs, and other kinds yet) are essentially about data: you save and retrieve groupings of elementary data (strings and numbers, mostly), with a lot of support for selecting and searching (a huge lot, for relational databases) and cross-language support. Databases have nothing to say about serializing Python objects into data, nor about deserializing Python objects back from data. The two approaches, databases and serialization, can even be used together. You can serialize Python objects into strings of bytes with pickle, and store those bytes using a databaseand vice versa at retrieval time. At a very elementary level, that's what the standard Python library shelve module does, for example, with pickle to serialize and deserialize and generally bsddb as the underlying simple database engine. So, don't think of the two approaches as being "in competition" with each otherrather, think of them as completing and complementing each other! For example, creating a database with an access method of db.DB_HASH, as shown in the recipe, may give maximum performance, but, as you'll have noticed when listing all records with the generator irecords that is also presented in the recipe, hashing puts records in apparently random, unpredictable order. If you need to access records in sorted order, you can use an access method of db.DB_BTREE instead. Berkeley DB also supports more advanced functionality, such as transactions, which you can enable through direct access but not via anydbm or shelve. For detailed documentation about all functionality of the Python Standard Library bsddb package, see http://pybsddb.sourceforge.net/bsddb3.html. For documentation, downloads, and more of the Berkeley DB itself, see http://www.sleepycat.com/. See Also Library Reference and Python in a Nutshell docs for modules anydbm, shelve, and bsddb; http://pybsddb.sourceforge.net/bsddb3.html for many more details about bsddb and bsddb3; http://www.sleepycat.com/ for downloads of, and very detailed documentation on, the Berkeley DB itself. Recipe 7.9. Accesssing a MySQL Database Credit: Mark Nenadov Problem You need to access a MySQL database. Solution The MySQLdb module makes this task extremely easy: import MySQLdb # Create a connection object, then use it to create a cursor con = MySQLdb.connect(host="127.0.0.1", port=3306, user="joe", passwd="egf42", db="tst") cursor = con.cursor( ) # Execute an SQL string sql = "SELECT * FROM Users" cursor.execute(sql) # Fetch all results from the cursor into a sequence and close the connection results = cursor.fetchall( ) con.close( ) Discussion MySQLdb is at http://sourceforge.net/projects/mysql-python. It is a plain and simple implementation of the Python DB API 2.0 that is suitable for Python 2.3 (and some older versions, too) and MySQL versions 3.22 to 4.0. MySQLdb, at the time of this writing, did not yet officially support Python 2.4. However, if you have the right C compiler installation to build Python extensions (as should be the case for all Linux, Mac OS X, and other Unix users, and many Windows developers), the current version of MySQLdb does in fact build from sources, install, and work just fine, with Python 2.4. A newer version of MySQLdb is in the works, with official support for Python 2.3 or later and MySQL 4.0 or later. As with all other Python DB API implementations (once you have downloaded and installed the needed Python extension and have the database engine it needs up and running), you start by importing the module and calling the connect function with suitable parameters. The keyword parameters you can pass when calling connect depend on the database involved: host (defaulting to the local host), user, passwd (password), and db (name of the database) are typical. In the recipe, I explicitly pass the default local host's IP address and the default MySQL port (3306), just to show that you can specify parameters explicitly even when you're passing their default values (e.g., to make your source code clearer and more readable and maintainable). The connect function returns a connection object, and you can proceed to call methods on this object; when you are done, call the close method. The method you most often call on a connection object is cursor, which returns a cursor object, which is what you use to send SQL commands to the database and fetch the commands' results. The underlying MySQL database engine does not in fact support SQL cursors, but that's no problemthe MySQLdb module emulates them on your behalf, quite transparently, for the limited cursor needs of the Python DB API 2.0.Of course, this doesn't mean that you can use SQL phrases like WHERE CURRENT OF CURSOR with a database that does not offer cursors! Once you have a cursor object in hand, you can call methods on it. The recipe uses the execute method to execute an SQL statement, and then the fetchall method to obtain all results as a sequence of tuplesone tuple per row in the result. You can use many refinements, but these basic elements of the Python DB API's functionality already suffice for many tasks. See Also The Python-MySQL interface module (http://sourceforge.net/projects/mysql-python); the Python DB API (http://www.python.org/topics/database/DatabaseAPI-2.0.html); DB API documentation in Python in a Nutshell. Recipe 7.10. Storing a BLOB in a MySQL Database Credit: Luther Blissett Problem You need to store a binary large object (BLOB) in a MySQL database. Solution The MySQLdb module does not support full-fledged placeholders, but you can make do with the module's escape_string function: import MySQLdb, cPickle # Connect to a DB, e.g., the test DB on your localhost, and get a cursor connection = MySQLdb.connect(db="test") cursor = connection.cursor( ) # Make a new table for experimentation cursor.execute("CREATE TABLE justatest (name TEXT, ablob BLOB)") try: # Prepare some BLOBs to insert in the table names = 'aramis', 'athos', 'porthos' data = { } for name in names: datum = list(name) datum.sort( ) data[name] = cPickle.dumps(datum, 2) # Perform the insertions sql = "INSERT INTO justatest VALUES(%s, %s)" for name in names: cursor.execute(sql, (name, MySQLdb.escape_string(data[name])) ) # Recover the data so you can check back sql = "SELECT name, ablob FROM justatest ORDER BY name" cursor.execute(sql) for name, blob in cursor.fetchall( ): print name, cPickle.loads(blob), cPickle.loads(data[name]) finally: # Done. Remove the table and close the connection. cursor.execute("DROP TABLE justatest") connection.close( ) Discussion MySQL supports binary data (BLOBs and variations thereof), but you should be careful when communicating such data via SQL. Specifically, when you use a normal INSERT SQL statement and need to have binary strings among the VALUES you're inserting, you have to escape some characters in the binary string according to MySQL's own rules. Fortunately, you don't have to figure out those rules for yourself: MySQL supplies a function that does the needed escaping, and MySQLdb exposes it to your Python programs as the escape_string function.This recipe shows a typical case: the BLOBs you're inserting come from cPickle.dumps, so they may represent almost arbitrary Python objects (although, in this case, we're just using them for a few lists of characters). The recipe is purely demonstrative and works by creating a table and dropping it at the end (using a try/finally statement to ensure that finalization is performed even if the program should terminate because of an uncaught exception). With recent versions of MySQL and MySQLdb, you don't even need to call the escape_string function anymore, so you can change the relevant statement to the simpler: cursor.execute(sql, (name, data[name])) See Also Recipe 7.11 and Recipe 7.12 for PostgreSQL-oriented and SQLite-oriented solutions to the same problem; the MySQL home page (http://www.mysql.org); the Python/MySQL interface module (http://sourceforge.net/projects/mysql-python). Recipe 7.11. Storing a BLOB in a PostgreSQL Database Credit: Luther Blissett Problem You need to store a BLOB in a PostgreSQL database. Solution PostgreSQL 7.2 and later supports large objects, and the psycopg module supplies a Binary escaping function: import psycopg, cPickle # Connect to a DB, e.g., the test DB on your localhost, and get a cursor connection = psycopg.connect("dbname=test") cursor = connection.cursor( ) # Make a new table for experimentation cursor.execute("CREATE TABLE justatest (name TEXT, ablob BYTEA)") try: # Prepare some BLOBs to insert in the table names = 'aramis', 'athos', 'porthos' data = { } for name in names: datum = list(name) datum.sort( ) data[name] = cPickle.dumps(datum, 2) # Perform the insertions sql = "INSERT INTO justatest VALUES(%s, %s)" for name in names: cursor.execute(sql, (name, psycopg.Binary(data[name])) ) # Recover the data so you can check back sql = "SELECT name, ablob FROM justatest ORDER BY name" cursor.execute(sql) for name, blob in cursor.fetchall( ): print name, cPickle.loads(blob), cPickle.loads(data[name]) finally: # Done. Remove the table and close the connection. cursor.execute("DROP TABLE justatest") connection.close( ) Discussion PostgreSQL supports binary data (BYTEA and variations thereof), but you should be careful when communicating such data via SQL. Specifically, when you use a normal INSERT SQL statement and need to have binary strings among the VALUES you're inserting, you have to escape some characters in the binary string according to PostgreSQL's own rules. Fortunately, you don't have to figure out those rules for yourself: PostgreSQL supplies functions that do all the needed escaping, and psycopg exposes such a function to your Python programs as the Binary function. This recipe shows a typical case: the BYTEAs you're inserting come from cPickle.dumps, so theymay represent almost arbitrary Python objects (although, in this case, we're just using them for a few lists of characters). The recipe is purely demonstrative and works by creating a table and dropping it at the end (using a TRy/finally statement to ensure finalization is performed even if the program should terminate because of an uncaught exception). Earlier PostgreSQL releases limited to a few kilobytes the amount of data you could store in a normal field of the database. To store really large objects, you had to use roundabout techniques to load the data into the database (such as PostgreSQL's nonstandard SQL function LO_IMPORT to load a data file as an object, which requires superuser privileges and data files that reside on the machine running the Postgre-SQL Server) and store a field of type OID in the table to be used later for indirect recovery of the data. Fortunately, none of these techniques are necessary anymore: since Release 7.1 (the current release at the time of writing is 8.0), PostgreSQL embodies the results of project TOAST, which removed the limitations on field-storage size and therefore the need for peculiar indirection. Module psycopg supplies the handy Binary function to let you escape any binary string of bytes into a form acceptable for placeholder substitution in INSERT and UPDATE SQL statements. See Also Recipe 7.10 and Recipe 7.12 for MySQL-oriented and SQLite-oriented solutions to the same problem; PostgresSQL's home page (http://www.postgresql.org/); the Python/PostgreSQL module (http://initd.org/software/psycopg). Recipe 7.12. Storing a BLOB in a SQLite Database Credit: John Barham Problem You need to store a BLOB in an SQLite database. Solution The PySQLite Python extension offers function sqlite.encode to let you insert binary strings in SQLite databases. You can also build a small adapter class based on that function: import sqlite, cPickle class Blob(object): ''' automatic converter for binary strings ''' def _ _init_ _(self, s): self.s = s def _quote(self): return "'%s'" % sqlite.encode(self.s) # make a test database in memory, get a cursor on it, and make a table connection = sqlite.connect(':memory:') cursor = connection.cursor( ) cursor.execute("CREATE TABLE justatest (name TEXT, ablob BLOB)") # Prepare some BLOBs to insert in the table names = 'aramis', 'athos', 'porthos' data = { } for name in names: datum = list(name) datum.sort( ) data[name] = cPickle.dumps(datum, 2) # Perform the insertions sql = 'INSERT INTO justatest VALUES(%s, %s)' for name in names: cursor.execute(sql, (name, Blob(data[name])) ) # Recover the data so you can check back sql = 'SELECT name, ablob FROM justatest ORDER BY name' cursor.execute(sql) for name, blob in cursor.fetchall( ): print name, cPickle.loads(blob), cPickle.loads(data[name]) # Done, close the connection (would be no big deal if you didn't, but...) connection.close( ) Discussion SQLite does not directly support binary data, but it still lets you declare such types for fields in a CREATE TABLE DDL statement. The PySQLite Python extension uses the declared types of fields to convert field values appropriately to Python values when you fetch data after an SQL SELECT from an SQLite database. However, you still need to be careful when communicating binary string data via SQL. Specifically, when you use INSERT or UPDATE SQL statements, and need to have binary stringsamong the VALUES you're passing, you need to escape some characters in the binary string according to SQLite's own rules. Fortunately, you don't have to figure out those rules for yourself: SQLite supplies the function to do the needed escaping, and PySQLite exposes that function to your Python programs as the sqlite.encode function. This recipe shows a typical case: the BLOBs you're inserting come from cPickle.dumps, so they may represent almost arbitrary Python objects (although, in this case, we're just using them for a few lists of characters). The recipe is purely demonstrative and works by creating a database in memory, so that the database is implicitly lost at the end of the script. While you could perfectly well call sqlite.encode directly on your binary strings at the time you pass them as parameters to a cursor's execute method, this recipe takes a slightly different tack, defining a Blob class to wrap binary strings and passing instances of that. When PySQLite receives as arguments instances of any class, the class must define a method named _quote, and PySQLite calls that method on each instance, expecting the method to return a string fully ready for insertion into an SQL statement. When you use this approach for more complicated classes of your own, you'll probably want to pass a decoders keyword argument to the connect method, to associate appropriate decoding functions to specific SQL types. By default, however, the BLOB SQL type is associated with the decoding function sqlite.decode, which is exactly the inverse of sqlite.encode; for the simple Blob class in this recipe, therefore, we do not need to specify any custom decoder, since the default one suits us perfectly well. See Also Recipe 7.10 and Recipe 7.11 for MySQL-oriented and PostgreSQL-oriented solutions to the same problem; SQLite's home page (http://www.sqlite.org/); the PySQLite manual (http://pysqlite.sourceforge.net/manual.html); the SQLite FAQ ("Does SQLite support a BLOB type?") at http://www.hwaci.com/sw/sqlite/faq.html#q12. Recipe 7.13. Generating a Dictionary Mapping Field Names to Column Numbers Credit: Thomas T. Jenkins Problem You want to access data fetched from a DB API cursor object, but you want to access the columns by field name, not by number. Solution Accessing columns within a set of database-fetched rows by column index is not very readable, nor is it robust should columns ever get reordered in a rework of the database's schema (a rare event, but it does occasionally happen). This recipe exploits the description attribute of Python DB API's cursor objects to build a dictionary that maps column names to index values, so you can use cursor_row[field_dict[fieldname]] to get the value of a named column: def fields(cursor): """ Given a DB API 2.0 cursor object that has been executed, returns a dictionary that maps each field name to a column index, 0 and up. """ results = { } for column, desc in enumerate(cursor.description): results[desc[0]] = column return results Discussion When you get a set of rows from a call to any of a cursor's various fetch . . . methods (fetchone, fetchmany, fetchall), it is often helpful to be able to access a specific column in a row by field name and not by column number. This recipe shows a function that takes a DB API 2.0 cursor object and returns a dictionary with column numbers keyed by field names. Here's a usage example (assuming you put this recipe's code in a module that you call dbutils.py somewhere on your Python sys.path). You must start with conn being a connection object for any DB API 2-compliant Python module. >>> c = conn.cursor( ) >>> c.execute('''select * from country_region_goal ... where crg_region_code is null''') >>> import pprint >>> pp = pprint.pprint >>> pp(c.description) (('CRG_ID', 4, None, None, 10, 0, 0), ('CRG_PROGRAM_ID', 4, None, None, 10, 0, 1), ('CRG_FISCAL_YEAR', 12, None, None, 4, 0, 1), ('CRG_REGION_CODE', 12, None, None, 3, 0, 1), ('CRG_COUNTRY_CODE', 12, None, None, 2, 0, 1), ('CRG_GOAL_CODE', 12, None, None, 2, 0, 1), ('CRG_FUNDING_AMOUNT', 8, None, None, 15, 0, 1))>>> import dbutils >>> field_dict = dbutils.fields(c) >>> pp(field_dict) {'CRG_COUNTRY_CODE': 4, 'CRG_FISCAL_YEAR': 2, 'CRG_FUNDING_AMOUNT': 6, 'CRG_GOAL_CODE': 5, 'CRG_ID': 0, 'CRG_PROGRAM_ID': 1, 'CRG_REGION_CODE': 3} >>> row = c.fetchone( ) >>> pp(row) (45, 3, '2000', None, 'HR', '26', 48509.0) >>> ctry_code = row[field_dict['CRG_COUNTRY_CODE']] >>> print ctry_code HR >>> fund = row[field_dict['CRG_FUNDING_AMOUNT']] >>> print fund 48509.0 If you find accesses such as row[field_dict['CRG_COUNTRY_CODE']] to be still inelegant, you may want to get fancier and wrap the row as well as the dictionary of fields into an object allowing more elegant accessa simple example might be: class neater(object): def _ _init_ _(self, row, field_dict): self.r = row self.d = field_dict def _ _getattr_ _(self, name): try: return self.r[self.d[name]] except LookupError: raise AttributeError If this neater class was also in your dubtils module, you could then continue the preceding interactive snippet with, for example: >>> row = dbutils.neater(row, field_dict) >>> print row.CRG_FUNDING_AMOUNT 48509.0 However, if you're tempted by such fancier approaches, I suggest that, rather than rolling your own, you have a look at the dbtuple module showcased in Recipe 7.14. Reusing good, solid, proven code is a much smarter approach than writing your own infrastructure. See Also Recipe 7.14 for a slicker and more elaborate approach to a very similar task, facilitated by reusing the third-party dbtuple module. Recipe 7.14. Using dtuple for Flexible Accessto Query Results Credit: Steve Holden, Hamish Lawson, Kevin Jacobs Problem You want flexible access to sequences, such as the rows in a database query, by either name or column number. Solution Rather than coding your own solution, it's often more clever to reuse a good existing one. For this recipe's task, a good existing solution is packaged in Greg Stein's dtuple module: import dtuple import mx.ODBC.Windows as odbc flist = ["Name", "Num", "LinkText"] descr = dtuple.TupleDescriptor([[n] for n in flist]) conn = odbc.connect("HoldenWebSQL") # Connect to a database curs = conn.cursor( ) # Create a cursor sql = """SELECT %s FROM StdPage WHERE PageSet='Std' AND Num<25 ORDER BY PageSet, Num""" % ", ".join(flist) print sql curs.execute(sql) rows = curs.fetchall( ) for row in rows: row = dtuple.DatabaseTuple(descr, row) print "Attribute: Name: %s Number: %d" % (row.Name, row.Num or 0) print "Subscript: Name: %s Number: %d" % (row[0], row[1] or 0) print "Mapping: Name: %s Number: %d" % (row["Name"], row["Num"] or 0) conn.close( ) Discussion Novice Python programmers are sometimes deterred from using databases because query results are presented by DB API-compliant modules as a list of tuples. Since tuples can only be numerically subscripted, code that uses the query results becomes opaque and difficult to maintain. Greg Stein's dtuple module, available from http://www.lyra.org/greg/python/dtuple.py, helps by defining two useful classes: TupleDescriptor and DatabaseTuple. To access an arbitrary SQL database, this recipe uses the ODBC protocol through the mxODBC module, http://www.egenix.com/files/python/mxODBC.html, but nothing relevant to the recipe's task would change if any other standard DB API-compliant module was used instead. The TupleDescriptor class creates a description of tuples from a list of sequences, the first element of each subsequence being a column name. It is often convenient to describe data with such sequences. For example, in an interactive forms-based application, each column name might be followed by validation parameters such as data type and allowable length.TupleDescriptor's purpose is to allow the creation of DatabaseTuple objects. In this particular application, no other information about the columns is needed beyond the names, so the required list of sequences is a list of singleton lists (meaning lists that have just one element each), constructed from a list of field names using a list comprehension. Created from TupleDescriptor and a tuple such as a database row, DatabaseTuple is an object whose elements can be accessed by numeric subscript (like a tuple) or column-name subscript (like a dictionary). If column names are legal Python names, you can also access the columns in your DatabaseTuple as attributes. A purist might object to this crossover between items and attributes, but it's a highly pragmatic choice in this case. Python is nothing if not a highly pragmatic language, so I see nothing wrong with this convenience. To demonstrate the utility of DatabaseTuple, the simple test program in this recipe creates a TupleDescriptor and uses it to convert each row retrieved from an SQL query into DatabaseTuple. Because the sample uses the same field list to build both TupleDescriptor and the SQL SELECT statement, it demonstrates how database code can be parameterized relatively easily. Alternatively, if you wish to get all the fields (an SQL SELECT * query), and dynamically get the field names from the cursor, as previously described in Recipe 7.13, you can do so. Just remove variable flist, which you don't need any more, and move the construction of variable descr to right after the call to the cursor's execute method, as follows: curs.execute(sql) descr = dtuple.TupleDescriptor(curs.description) The rest of the recipe can remain unchanged. A more sophisticated approach, with functionality similar to dtuple's and even better performance, is offered by the Python Database Row Module (also known as db_row) made freely available by the OPAL Group. For downloads and information, visit http://opensource.theopalgroup.com/. Module pysqlite, which handles relational databases in memory or in files by wrapping the SQLite library, does not return real tuples from such methods as fetchall: rather, it returns instances of a convenience class that wraps tuple and also allows field access with attribute- access syntax, much like the approaches mentioned in this recipe. See Also Recipe 7.13 for a simpler, less functionally rich way to convert field names to column numbers; the dtuple module is at http://www.lyra.org/greg/python/dtuple.py; OPAL's db_row is at http://opensource.theopalgroup.com/; SQLite, a fast, lightweight, embedded relational database (http://www.sqlite.org/), and its Python DB API interface module pysqlite (http://pysqlite.sourceforge.net/). Recipe 7.15. Pretty-Printing the Contents of Database Cursors Credit: Steve Holden, Farhad Fouladi, Rosendo Martinez, David Berry, Kevin Ryan Problem You want to present a query's result with appropriate column headers (and optionally widths), but you do not want to hard-code this information in your program. Indeed, you may not even know the column headers and widths at the time you're writing the code. Solution Discovering the column headers and widths dynamically is the most flexible approach, and it gives you code that's highly reusable over many such presentation tasks: def pp(cursor, data=None, check_row_lengths=False): if not data: data = cursor.fetchall( ) names = [ ] lengths = [ ] rules = [ ] for col, field_description in enumerate(cursor.description): field_name = field_description[0] names.append(field_name) field_length = field_description[2] or 12 field_length = max(field_length, len(field_name)) if check_row_lengths: # double-check field length, if it's unreliable data_length = max([ len(str(row[col])) for row in data ]) field_length = max(field_length, data_length) lengths.append(field_length) rules.append('-' * field_length) format = " ".join(["%%-%ss" % l for l in lengths]) result = [ format % tuple(names), format % tuple(rules) ] for row in data: result.append(format % tuple(row)) return "\n".join(result) Discussion Relational databases are often perceived as difficult to use. The Python DB API can make them much easier to use, but if your programs work with several different DB engines, it's sometimes tedious to reconcile the implementation differences between the various modules, and, even more, between the engines they connect to. One of the problems of dealing with databases is presenting the result of a query when you may not know much about the data. This recipe uses the cursor's description attribute to try to provide appropriate headings. The recipe optionally examines each output row to ensure that column widths are adequate. In some cases, a cursor can yield a solid description of the data it returns, but not all databasemodules are kind enough to supply cursors that do so. The pretty printer function pp shown in this recipe's Solution takes as its first argument a cursor, on which you have just executed a retrieval operation (generally the execute of an SQL SELECT statement). It also takes an optional argument for the returned data; to use the data for other purposes, retrieve the data from the cursor, typically with fetchall, and pass it in as pp's data argument. The second optional argument tells the pretty printer to determine the column lengths from the data, rather than trusting the cursor's description; checking the data for column lengths can be time-consuming, but is helpful with some RDBMS engines and DB API module combinations, where the widths given by the cursor's description attribute can be inaccurate. A simple test program shows the value of the second optional argument when a Microsoft Jet database is used through the mxODBC module: import mx.ODBC.Windows as odbc import dbcp # contains pp function conn = odbc.connect("MyDSN") curs = conn.cursor( ) curs.execute("""SELECT Name, LinkText, Pageset FROM StdPage ORDER BY PageSet, Name""") rows = curs.fetchall( ) print "\n\nWithout rowlens:" print dbcp.pp(curs, rows) print "\n\nWith rowlens:" print dbcp.pp(curs, rows, rowlens=1) conn.close( ) In this case, the cursor's description does not include column lengths. The first output shows that the default column length of 12 is too short. The second output corrects this by examining the data: Without rowlens: Name LinkText Pageset ------------ ------------ ------------ ERROR ERROR: Cannot Locate Page None home Home None consult Consulting Activity Std ffx FactFaxer Std hardware Hardware Platforms Std python Python Std rates Rates Std technol Technologies Std wcb WebCallback Std With rowlens: Name LinkText Pageset ------------ ------------------------- ------------ ERROR ERROR: Cannot Locate Page None home Home None consult Consulting Activity Std ffx FactFaxer Std hardware Hardware Platforms Std python Python Std rates Rates Std technol Technologies Std wcb WebCallback Std Module pysqlite, which handles relational databases in memory or in files by wrapping the SQLite library, is another example of a DB API module whose cursors' descriptions do not contain reliable values for field lengths. Moreover, pysqlite does not return real tuples from suchmethods as fetchall: rather, it returns instances of a convenience class which wraps tuple and also allocws field access with attribute access syntax, much like the approaches presented in Recipe 7.14. To deal with such small variations from the DB API specifications, this recipe carefully uses tuple(row), not just row, as the right-hand operand of operator % in the statement result.append(format % tuple(row)). Python's semantics specify that if the right- hand operand is not a tuple, then the left-hand (format string) operand may contain only one format specifier. This recipe uses a tuple as the right-hand operand because the whole point of the recipe is to build and use a format string with many format specifiers, one per field. This recipe's function is useful during testing, since it lets you easily verify that you are indeed retrieving what you expect from the database. The output is pretty enough to display ad hoc query outputs to users. The function currently makes no attempt to represent null values other than the None the DB API returns, though it could easily be modified to show a null string or some other significant value. See Also The mxODBC package, a DB API-compatible interface to ODBC (http://www.egenix.com/files/python/mxODBC.html); SQLite, a fast, lightweight embedded relational database (http://www.sqlite.org/), and its Python DB API interface module pysqlite (http://pysqlite.sourceforge.net/). Recipe 7.16. Using a Single Parameter-Passing Style Across Various DB API Modules Credit: Denis S. Otkidach Problem You want to write Python code that runs under any DB API compliant module, but such modules can use different styles to allow parameter passing. Solution We need a set of supporting functions to convert SQL queries and parameters to any of the five possible parameter-passing styles: class Param(object): ''' a class to wrap any single parameter ''' def _ _init_ _(self, value): self.value = value def _ _repr_ _(self): return 'Param(%r)' % (self.value,) def to_qmark(chunks): ''' prepare SQL query in '?' style ''' query_parts = [ ] params = [ ] for chunk in chunks: if isinstance(chunk, Param): params.append(chunk.value) query_parts.append('?') else: query_parts.append(chunk) return ''.join(query_parts), params def to_numeric(chunks): ''' prepare SQL query in ':1' style ''' query_parts = [ ] params = [ ] for chunk in chunks: if isinstance(chunk, Param): params.append(chunk.value) query_parts.append(':%d' % len(params)) else: query_parts.append(chunk) # DCOracle2 needs, specifically, a _tuple_ of parameters: return ''.join(query_parts), tuple(params) def to_named(chunks): ''' prepare SQL query in ':name' style ''' query_parts = [ ] params = { } for chunk in chunks: if isinstance(chunk, Param): name = 'p%d' % len(params) params[name] = chunk.value query_parts.append(':%s' % name) else: query_parts.append(chunk) return ''.join(query_parts), params def to_format(chunks): ''' prepare SQL query in '%s' style ''' query_parts = [ ] params = [ ] for chunk in chunks: if isinstance(chunk, Param): params.append(chunk.value) query_parts.append('%s') else: query_parts.append(chunk.replace('%', '%%')) return ''.join(query_parts), params def to_pyformat(chunks): ''' prepare SQL query in '%(name)s' style ''' query_parts = [ ] params = { } for chunk in chunks: if isinstance(chunk, Param): name = 'p%d' % len(params) params[name] = chunk.value query_parts.append('%%(%s)s' % name) else: query_parts.append(chunk.replace('%', '%%')) return ''.join(query_parts), params converter = { } for paramstyle in ('qmark', 'numeric', 'named', 'format', 'pyformat'): converter[paramstyle] = globals['to_' + param_style] def execute(cursor, converter, chunked_query): query, params = converter(chunked_query) return cursor.execute(query, params) if _ _name_ _=='_ _main_ _': query = ('SELECT * FROM test WHERE field1>', Param(10), ' AND field2 LIKE ', Param('%value%')) print 'Query:', query for paramstyle in ('qmark', 'numeric', 'named', 'format', 'pyformat'): print '%s: %r' % (paramstyle, converter[param_style](query)) Discussion The DB API specification is quite handy, but it has one most annoying problem: it allows compliant modules to use any of five parameter styles. So you cannot necessarily switch to another database just by changing the database module: if the parameter-passing styles of two such modules differ, you need to rewrite all SQL queries that use parameter substitution. Using this recipe, you can improve this situation a little. Pick the appropriate converter from the converter dictionary (indexing it with the paramstyle attribute of your current DB API module), write your queries as mixed chunks of SQL strings and instances of the provided Param class (as exemplified in the if _ _name_ _=='_ _main_ _' part of the recipe), and execute your queries through the execute function in this recipe. Not a perfectly satisfactory solution, by any means, but way better than nothing! See Also The DB API docs at http://www.python.org/peps/pep-0249.html; the list of DB API-compliantmodules at http://www.python.org/topics/database/modules.html. Recipe 7.17. Using Microsoft Jet via ADO Credit: Souman Deb Problem You need to access a Microsoft Jet database via Microsoft's ADO, for example from a Python- coded CGI script for the Apache web-server. Solution The CGI script must live in Apache's cgi-bin directory and can use the PyWin32 extensions to connect, via COM, to ADO and hence to Microsoft Jet. For example: #!C:\Python23\python print "Content-type:text/html\n\n" import win32com db='C:\\Program Files\\Microsoft Office\\Office\\Samples\\Northwind.mdb' MAX_ROWS=2155 def connect(query): con = win32com.client.Dispatch('ADODB.Connection') con.Open("Provider=Microsoft.Jet.OLEDB.4.0; Data Source="+db) result_set = con.Execute(query + ';') con.Close( ) return result_set def display(columns, MAX_ROWS): print "" print "" print "" print "" print "" print "" for k in range(MAX_ROWS): print "" for field in columns: print "" print "" print "
    Order IDProductUnit PriceQuantityDiscount
    ", field[k], "
    " result_set = connect("select * from [Order details]") columns = result_set[0].GetRows(MAX_ROWS) display(columns, MAX_ROWS) result_set[0].Close Discussion This recipe uses the "Northwind Database" example that Microsoft distributes with several of its products, such as Microsoft Access. To run this recipe, you need a machine running Microsoft Windows with working installations of other Microsoft add-ons such as OLEDB, ADO, and the Jet database driver, which is often (though not correctly) known as "the Access database". (Microsoft Access is a product to build database frontend applications, and it can work with otherdatabase drivers, such as Microsoft SQL Server, not just with the freely distributable and downloadable Microsoft Jet database drivers.) Moreover, you need an installation of Mark Hammond's PyWin32 package (formerly known as win32all); the Python distribution known as ActivePython, from ActiveState, comes with (among other things) PyWin32 already installed. If you want to run this recipe specifically as an Apache CGI script, of course, you also need to install Apache and to place this script in the cgi-bin directory where Apache expects to find CGI scripts (the location of the cgi-bin directory depends on how you have installed Apache on your machine). Make sure that the paths in the script are correct, depending on where, on your machine, you have installed the python.exe file you want to use, and the Northwind.mdb database you want to query. The paths indicated in the recipe correspond to default installations of Python 2.3 and the "Northwind" example database. If the script doesn't work correctly, check the Apache error.log file, where you will find error messages that may help you find out what kind of error you're dealing with. To try the script, assuming that, for example, you have saved it as cgi-bin/adoexample.py and that your Apache server is running correctly, visit with any browser the URL http://localhost/cgi- bin/adoexample.py. One known limitation of the interface between Python and Jet databases with ADO is on fields of type currency: such fields are returned as some strange tuples, rather than as plain numbers. This recipe does not deal with that limitation. See Also Documentation for the Win32 API in PyWin32 (http://starship.python.net/crew/mhammond/win32/Downloads.html) or ActivePython (http://www.activestate.com/ActivePython/); Windows API documentation available from Microsoft (http://msdn.microsoft.com); Mark Hammond and Andy Robinson, Python Programming on Win32 (O'Reilly). Recipe 7.18. Accessing a JDBC Database from a Jython Servlet Credit: Brian Zhou Problem You're writing a servlet in Jython, and you need to connect to a database server (such as Oracle, Sybase, Microsoft SQL Server, or MySQL) via JDBC. Solution The technique is basically the same for any kind of database, give or take a couple of statements. Here's the code for when your database is Oracle: import java, javax class emp(javax.servlet.http.HttpServlet): def doGet(self, request, response): ''' a Servlet answers a Get query by writing to the response's output stream. In this case we ignore the request, though in normal, non-toy cases that's where we get form input from. ''' # we answer in plain text, so set the content type accordingly response.setContentType("text/plain") # get the output stream, use it for the query, then close it out = response.getOutputStream( ) self.dbQuery(out) out.close( ) def dbQuery(self, out): # connect to the Oracle driver, building an instance of it driver = "oracle.jdbc.driver.OracleDriver" java.lang.Class.forName(driver).newInstance( ) # get a connection to the Oracle driver w/given user and password server, db = "server", "ORCL" url = "jdbc:oracle:thin:@" + server + ":" + db usr, passwd = "scott", "tiger" conn = java.sql.DriverManager.getConnection(url, usr, passwd) # send an SQL query to the connection query = "SELECT EMPNO, ENAME, JOB FROM EMP" stmt = conn.createStatement( ) if stmt.execute(query): # get query results and print the out to the out stream rs = stmt.getResultSet( ) while rs and rs.next( ): out.println(rs.getString("EMPNO")) out.println(rs.getString("ENAME")) out.println(rs.getString("JOB")) out.println( ) stmt.close( ) conn.close( )When your database is Sybase or Microsoft SQL Server, use the following (we won't repeat the comments from the preceding Oracle example, since they apply identically here): import java, javax class titles(javax.servlet.http.HttpServlet): def doGet(self, request, response): response.setContentType("text/plain") out = response.getOutputStream( ) self.dbQuery(out) out.close( ) def dbQuery(self, out): driver = "sun.jdbc.odbc.JdbcOdbcDriver" java.lang.Class.forName(driver).newInstance( ) # Use "pubs" DB for mssql and "pubs2" for Sybase url = "jdbc:odbc:myDataSource" usr, passwd = "sa", "password" conn = java.sql.DriverManager.getConnection(url, usr, passwd) query = "select title, price, ytd_sales, pubdate from titles" stmt = conn.createStatement( ) if stmt.execute(query): rs = stmt.getResultSet( ) while rs and rs.next( ): out.println(rs.getString("title")) if rs.getObject("price"): out.println("%2.2f" % rs.getFloat("price")) else: out.println("null") if rs.getObject("ytd_sales"): out.println(rs.getInt("ytd_sales")) else: out.println("null") out.println(rs.getTimestamp("pubdate").toString( )) out.println( ) stmt.close( ) conn.close( ) And here's the code for when your database is MySQL: import java, javax class goosebumps(javax.servlet.http.HttpServlet): def doGet(self, request, response): response.setContentType("text/plain") out = response.getOutputStream( ) self.dbQuery(out) out.close( ) def dbQuery(self, out): driver = "org.gjt.mm.mysql.Driver" java.lang.Class.forName(driver).newInstance( ) server, db = "server", "test" usr, passwd = "root", "password" url = "jdbc:mysql://%s/%s?user=%s&password=%s" % ( server, db, usr, passwd) conn = java.sql.DriverManager.getConnection(url) query = "select country, monster from goosebumps" stmt = conn.createStatement( ) if stmt.execute(query): rs = stmt.getResultSet( ) while rs and rs.next( ): out.println(rs.getString("country")) out.println(rs.getString("monster")) out.println( ) stmt.close( ) Discussion You might want to use different JDBC drivers and URLs, but you can see that the basic technique is quite simple and straightforward. This recipe's code uses a content type of text/plain because the recipe is about accessing the database, not about formatting the data you get from it. Obviously, you can change this content type to whichever is appropriate for your application. In each case, the basic technique is first to instantiate the needed driver (whose package name, as a string, we place in variable driver) via the Java dynamic loading facility. The forName method of the java.lang.Class class loads and provides the relevant Java class, and that class' newInstance method ensures that the driver we need is instantiated. Then, we can call the getConnection method of java.sql.DriverManager with the appropriate URL (or username and password, where needed) and thus obtain a connection object to place in the conn variable. From the connection object, we can create a statement object with the createStatement method and use it to execute a query that we have in the query string variable with the execute method. If the query succeeds, we can obtain the results with the geTResultSet method. Finally, Oracle and MySQL allow easy sequential navigation of the result set to present all results, while Sybase and Microsoft SQL Server need a bit more care. Overall, the procedure is similar in all cases. See Also The Jython site (http://www.jython.org); JDBC's home page (http://java.sun.com/products/jdbc). Recipe 7.19. Using ODBC to Get Excel Data with Jython Credit: Zabil CM Problem Your Jython script needs to extract data from a Microsoft Excel file. Solution Jython, just like Java, can access ODBC through the JDBC-ODBC Bridge, and Microsoft Excel can in turn be queried via ODBC: from java import lang, sql lang.Class.forName('sun.jdbc.odbc.JdbcOdbcDriver') excel_file = 'values.xls' connection = sql.DriverManager.getConnection( 'jdbc:odbc:Driver={Microsoft Excel Driver (*.xls)};DBQ=%s;READONLY=true}' % excelfile, '', '') # Sheet1 is the name of the Excel workbook we want. The field names for the # query are implicitly set by the values for each column in the first row. record_set = connection.createStatement( ).executeQuery( 'SELECT * FROM [Sheet1$]') # print the first-column field of every record (==row) while record_set.next( ): print record_set.getString(1) # we're done, close the connection and recordset record_set.close( ) connection.close( ) Discussion This recipe is most easily used on Microsoft Windows, where installing and configuring ODBC, and the Microsoft Excel ODBC driver in particular, is best supported. However, with suitable commercial products, you can equally well use the recipe on an Apple Macintosh or just about any other Unix version on the planet. Using ODBC rather than alternate ways to access Microsoft Excel has one substantial advantage that is not displayed in this recipe: with ODBC, you can use a broad subset of SQL. For example, you can easily extract a subset of a workbook's row by adding a WHERE clause, such as: SELECT * FROM [Sheet1$] WHERE DEPARTMENT=9 Since all of the selection logic can be easily expressed in the SQL string you pass to the executeQuery method, this approach lends itself particularly well to being encapsulated in a simple reusable function. If you're coding in Classic Python (CPython) rather than Jython, you can't use JDBC, but you can use ODBC directly (typically in the DB API-compliant way supported by mxODBC, http://www.egenix.com/files/python/mxODBC.html) to perform this recipe's task in a similarway. See Also The Jython site (http://www.jython.org); JDBC's home page (http://java.sun.com/products/jdbc); Recipe 12.7, for another way to access Excel data (by parsing the XML file that Excel can be asked to output). Chapter 8. Debugging and Testing Introduction Recipe 8.1. Disabling Execution of Some Conditionals and Loops Recipe 8.2. Measuring Memory Usage on Linux Recipe 8.3. Debugging the Garbage-Collection Process Recipe 8.4. Trapping and Recording Exceptions Recipe 8.5. Tracing Expressions and Comments in Debug Mode Recipe 8.6. Getting More Information from Tracebacks Recipe 8.7. Starting the Debugger Automatically After an Uncaught Exception Recipe 8.8. Running Unit Tests Most Simply Recipe 8.9. Running Unit Tests Automatically Recipe 8.10. Using doctest with unittest in Python 2.4 Recipe 8.11. Checking Values Against Intervals in Unit Testing Introduction Credit: Mark Hammond, co-author of Python Programming on Win32 (O'Reilly) The first computer I had in my home was a 64 KB Z80 CP/M machine. Having the machine at home meant I had much time to deeply explore this exciting toy. Turbo Pascal had just been released, and it seemed the obvious progression from the various BASIC dialects and assemblers I had been using. Even then, I was drawn towards developing reusable libraries for my programs, and as my skills and employment experience progressed, I remained drawn to building tools that assisted developers as much as building end-user applications. Building tools for developers means that debugging and testing are often in the foreground. Although images of an interactive debugger may pop into your head, the concepts of debugging and testing are much broader than you may initially think. Debugging and testing are sometimes an inseparable cycle. Testing will often lead to the discovery of bugs. You debug until you believe you understand the cause of the error and make the necessary changes. Rinse and repeat as required. Debugging and testing often are more insidious. I am a big fan of Python's assert statement, and every time I use it, I am debugging and testing my program. Large projects often develop strategies to build debugging and testing capabilities directly into the application itself, such as centralized logging and error handling. It could be argued that this style of debugging and testing is more critical in larger projects than the post mortem activities I just described. Python, in particular, supports a variety of techniques to help developers in their endeavors. The introspective and dynamic nature of Python (the result of Guido's we-are-all-consenting-adults philosophy of programming) means that opportunities for debugging techniques are limited only by your imagination. You can replace functions at runtime, add methods to classes, and extract everything about your program that there is to know. All at runtime, and all quite simple and Pythonic. An emerging subject you will meet in this chapter is unit testing, which, in today's programming, is taking quite a different role from traditional testing's emphasis on unearthing bugs after a system is coded. Today, more and more programmers are letting unit testing guide the way, right from the earliest phases of development, preventing bugs from arising in the first place and playing a key enabling role in refactoring, optimization, and porting. Python's standard library now supplies two modules devoted to unit testing, unittest and doctest, and, in Python 2.4, a bridge between them, which you'll see highlighted in one of this chapter's recipes. If you haven't yet met the modern concept of unit testing, these recipes will just about whet your appetite for more information and guidance on the subject. Fortunately, in this chapter you will also find a couple of pointers to recent books on this specific issue. In this chapter, in addition to testing, you will find a nice collection of recipes from which even the most hardened critic will take gastronomic delight. Whether you want customized error logging, deep diagnostic information in Python tracebacks, or even help with your garbage, you have come to the right place. So tuck in your napkin; your next course has arrived! Recipe 8.1. Disabling Execution of Some Conditionals and Loops Credit: Chris McDonough, Srinivas B, Dinu Gherman Problem While developing or debugging, you want certain conditional or looping sections of code to be temporarily omitted from execution. Solution The simplest approach is to edit your code, inserting 0: # right after the if or while keyword. Since 0 evaluates as false, that section of code will not execute. For example: if i < 1: doSomething( ) while j < k: j = fleep(j, k) into: if 0: # i < 1: doSomething( ) while 0: # j < k: j = fleep(j, k) If you have many such sections that must simultaneously switch on and off during your development and debug sessions, an alternative is to define a boolean variable (commonly known as a flag), say doit = False, and code: if doit and i < 1: doSomething( ) while doit and j < k: j = fleep(j, k) This way, you can temporarily switch the various sections on again by just changing the flag setting to doit = True, and easily flip back and forth. You can even have multiple such flags. Do remember to remove the doit and parts once you're done developing and debugging, since at that point all they would do is slow things down. Discussion Of course, you have other alternatives, too. Most good editors have commands to insert or remove comment markers from the start of each line in a marked section, like Alt-3 and Alt-4 in the editor of the IDLE IDE (Integrated Development Environment) that comes with Python; a common convention in such editors is to start such temporarily commented-out lines with twocomment markers, ##, to distinguish them from "normal" comments. One Python-specific technique you can use is the _ _debug_ _ read-only global boolean variable. _ _debug_ _ is true when Python is running without the -O (optimize) command-line option, False when Python is running with that option. Moreover, the Python compiler knows about _ _debug_ _ and can completely remove any block guarded by if _ _debug_ _ when Python is running with the command-line optimization option, thus saving memory as well as execution time. See Also The section on the _ _debug_ _ flag and the assert statement in the Language Reference and Python in a Nutshell. Recipe 8.2. Measuring Memory Usage on Linux Credit: Jean Brouwers Problem You need to monitor how much memory your Python application, running under Linux, is currently using. However, the standard library module resource does not work correctly on Linux. Solution We can code our own resource measurements based on Linux's /proc pseudo-filesystem: import os _proc_status = '/proc/%d/status' % os.getpid( ) _scale = {'kB': 1024.0, 'mB': 1024.0*1024.0, 'KB': 1024.0, 'MB': 1024.0*1024.0} def _VmB(VmKey): ''' given a VmKey string, returns a number of bytes. ''' # get pseudo file /proc//status try: t = open(_proc_status) v = t.read( ) t.close( ) except IOError: return 0.0 # non-Linux? # get VmKey line e.g. 'VmRSS: 9999 kB\n ...' i = v.index(VmKey) v = v[i:].split(None, 3) # split on runs of whitespace if len(v) < 3: return 0.0 # invalid format? # convert Vm value to bytes return float(v[1]) * _scale[v[2]] def memory(since=0.0): ''' Return virtual memory usage in bytes. ''' return _VmB('VmSize:') - since def resident(since=0.0): ''' Return resident memory usage in bytes. ''' return _VmB('VmRSS:') - since def stacksize(since=0.0): ''' Return stack size in bytes. ''' return _VmB('VmStk:') - since Discussion Each of the functions in this recipe takes an optional argument since because the typical usage of these functions is to find out how much more memory (virtual, resident, or stack) has been used due to a certain section of code. Having since as an optional argument makes this typical usage quite simple and elegant: m0 = memory( ) section of code you're monitoring m1 = memory(m0) print 'The monitored section consumed', m1, 'bytes of virtual memory'. Getting and parsing the contents of pseudo-file /proc/pid/status is probably not the most efficient way to get data about memory usage, and it is not portable to non-Linux systems. However, it is a very simple and easy-to-code approach, and after all, on a non-Linux Unix system, you can use the resource module from the Python Standard Library. In fact, you can use resource on Linux, but the various fields relevant to memory consumption, such as ru_maxrss, all have a constant value of 0, just like the various memory-consumption fields in the output of the time shell command under Linux. The root cause of this situation is a limitation in the Linux implementation of the geTRusage system call, documented in man getrusage. See Also Documentation on the resource standard library module in the Library Reference. Recipe 8.3. Debugging the Garbage-Collection Process Credit: Dirk Holtwick Problem You know that memory is leaking from your program, but you have no indication of what exactly is being leaked. You need more information to help you figure out where the leaks are coming from, so you can remove them and lighten the garbage-collection work periodically performed by the standard gc module. Solution The gc module lets you dig into garbage-collection issues: import gc def dump_garbage( ): """ show us what the garbage is about """ # Force collection print "\nGARBAGE:" gc.collect( ) print "\nGARBAGE OBJECTS:" for x in gc.garbage: s = str(x) if len(s) > 80: s = s[:77]+'...' print type(x),"\n ", s if _ _name_ _=="_ _main_ _": gc.enable( ) gc.set_debug(gc.DEBUG_LEAK) # Simulate a leak (a list referring to itself) and show it l = [ ] l.append(l) del l dump_garbage( ) # emits: # GARBAGE: # gc: collectable # GARBAGE OBJECTS: # # [[...]] Discussion In addition to the normal debugging output of gc, this recipe shows the garbage objects, to help you get an idea of where the leak may be. Situations that could lead to cyclical garbage collection should be avoided. Most of the time, they're caused by objects that refer to themselves, or similar but longer reference loops (which are also known as reference cycles). Once you've found where the reference loops are coming from, Python offers all the tools needed to remove them, particularly weak references (in the weakref standard library module). Butespecially in big programs, you first have to get an idea of where to find the leak before you can remove it and enhance your program's performance. For this purpose, it's good to know what the objects being leaked contain, and the dump_garbage function in this recipe comes in quite handy on such occasions. This recipe works by first calling gc.set_debug to tell the gc module to keep the leaked objects in its gc.garbage list rather than recycling them. Then, this recipe's dump_garbage function calls gc.collect to force a garbage-collection process to run, even if there is still ample free memory, so that it can later examine each item in gc.garbage and print out its type and contents (limiting the printout to no more than 80 characters per garbage object, to avoid flooding the screen with huge chunks of information). See Also Documentation for the gc and weakref modules in the Library Reference and Python in a Nutshell. Recipe 8.4. Trapping and Recording Exceptions Credit: Mike Foord Problem You need to trap exceptions, record their tracebacks and error messages, and then proceed with the execution of your program. Solution A typical case is a program that processes many independent files one after the other. Some files may be malformed and cause exceptions. You need to trap such exceptions, record the error information, then move on to process subsequent files. For example: import cStringIO, traceback def process_all_files(all_filenames, fatal_exceptions=(KeyboardInterrupt, MemoryError) ): bad_filenames = { } for one_filename in all_filenames: try: process_one_file(one_filename): except fatal_exceptions: raise except Exception: f = cStringIO.StringIO( ) traceback.print_exc(file=f) bad_filenames[one_filename] = f.getvalue( ) return bad_filenames Discussion Because Python exceptions are very powerful tools, you need a clear and simple strategy to deal with them. This recipe will probably not fit your needs exactly, but it may be a good starting point from which to develop the right strategy for your applications. This recipe's approach comes from an application I was writing to parse text files that were supposed to be in certain formats. Malformed files could easily cause exceptions, and I needed to get those errors' tracebacks and messages to either fix my code to be more forgiving or fix malformed files; however, I also needed program execution to continue on subsequent files. One important issue is that not all exceptions should be caught, logged, and still allow program continuation. A KeyboardInterrupt exception means the user is banging on Ctrl-C (or Ctrl- Break, or some other key combination), specifically asking for your application to stop; we should, of course, honor such requests, not ignore them. A MemoryError means you have run out of memoryunless you've got huge caches of previous results that you can immediately delete to make more memory available, generally you can't do much about such a situation. Depending on your application and exact circumstances, other errors might well also be deemed just as fatal. So, process_all_files accepts a fatal_exceptions argument, a tuple of exception classes it should not catch (but which it should rather propagate), defaulting to the pair of exceptionclasses we just mentioned. The try/except statement is carefully structured to catch, and re- raise, any exception in those classes, with precedence over the general except Exception handler clause, which catches everything else. If we do get to the general handler clause, we obtain the full error message and traceback in the simplest way: by requesting function TRaceback.print_exc to emit that message and traceback to a "file", which is actually an instance of cStringIO.StringIO, a "file"-like object specifically designed to ease in-memory capture of information from functions that normally write to files. The getvalue method of the StringIO instance provides the message and traceback as a string, and we store the string in dictionary bad_filenames, using, as the corresponding key, the filename that appears to have caused the probl7em. process_all_files' for loop then moves on to the next file it must process. Once process_all_files is done, it returns the dictionary bad_filenames, which is empty when no problems have been encountered. Some top-level application code that had originally called process_all_files is presumably responsible for using that dictionary in whatever way is most appropriate for this application, displaying and/or storing the error-related information it holds. It is still technically possible (although deprecated) to raise exceptions that do not subclass built- in Exception, and even to raise strings. If you need to catch such totally anomalous cases (whose possibility will no doubt stay around for years for backwards compatibility), you need to add one last unconditional except clause to your try/except statement: except fatal_exceptions: raise except Exception: ... except: ... Of course, if what you want to do for all normal (nonfatal) exceptions, and for the weird anomalous cases, is exactly the same, you don't need a separate except Exception clausejust the unconditional except clause will do. However, you may normally want to log the occurrence of the weird anomalous cases in some different and more prominent way, because, these days (well into the twenty-first century), they're definitely not expected under any circumstance whatsoever. See Also Documentation for the standard modules TRaceback and cStringIO in the Library Reference and Python in a Nutshell; documentation for try/except and exception classes in the Language Reference and Python in a Nutshell. Recipe 8.5. Tracing Expressions and Comments in Debug Mode Credit: Olivier Dagenais Problem You are coding a program that cannot use an interactive, step-by-step debugger. Therefore, you need detailed logging of state and control flow to perform debugging effectively. Solution The extract_stack function from the TRaceback module is the key here because it lets your debugging code easily perform runtime introspection to find out about the code that called it: import sys, traceback traceOutput = sys.stdout watchOutput = sys.stdout rawOutput = sys.stdout # calling 'watch(secretOfUniverse)' prints out something like: # File "trace.py", line 57, in _ _testTrace # secretOfUniverse = 42 watch_format = ('File "%(fileName)s", line %(lineNumber)d, in' ' %(methodName)s\n %(varName)s <%(varType)s>' ' = %(value)s\n\n') def watch(variableName): if _ _debug_ _: stack = traceback.extract_stack( )[-2:][0] actualCall = stack[3] if actualCall is None: actualCall = "watch([unknown])" left = actualCall.find('(') right = actualCall.rfind(')') paramDict = dict(varName=actualCall[left+1:right]).strip( ), varType=str(type(variableName))[7:-2], value=repr(variableName), methodName=stack[2], lineNumber=stack[1], fileName=stack[0]) watchOutput.write(watch_format % paramDict) # calling 'trace("this line was executed")' prints out something like: # File "trace.py", line 64, in ? # this line was executed trace_format = ('File "%(fileName)s", line %(lineNumber)d, in' ' %(methodName)s\n %(text)s\n\n') def trace(text): if _ _debug_ _: stack = traceback.extract_stack( )[-2:][0] paramDict = dict(text=text, methodName=stack[2], lineNumber=stack[1], fileName=stack[0]) watchOutput.write(trace_format % paramDict) # calling 'raw("some raw text")' prints out something like: # Just some raw text def raw(text): if _ _debug_ _: rawOutput.write(text) Discussion Many of the different kinds of programs one writes today don't make it easy to use traditional, interactive step-by-step debuggers. Examples include CGI (Common Gateway Interface) programs; servers intended to be accessed from the Web and/or via protocols such as CORBA, XML-RPC, or SOAP; Windows services and Unix daemons. You can remedy this lack of interactive debugging by sprinkling a bunch of print statements all through the program, but this approach is unsystematic and requires cleanup when a given problem is fixed. This recipe shows that a better-organized approach is quite feasible, by supplying a few functions that allow you to output the value of an expression, a variable, or a function call, with scope information, trace statements, and general comments. The key is the extract_stack function from the TRaceback module. traceback.extract_stack returns a list of tuples with four itemsproviding the filename, line number, function name, and source code of the calling statementfor each call in the stack. Item [-2] (the penultimate item) of this list is the tuple of information about our direct caller, and that's the one we use in this recipe to prepare the information to emit on file-like objects bound to the traceOutput and watchOutput variables. If you bind the traceOutput, watchOutput, or rawOutput variables to an appropriate file-like object, each kind of output is redirected appropriately. When _ _debug_ _ is false (i.e., when you run the Python interpreter with the -O or -OO switch), all the debugging-related code is automatically eliminated. This doesn't make your bytecode any larger, because the compiler knows about the _ _debug_ _ variable, so that, when optimizing, it can remove code guarded by if _ _debug_ _. Here is a usage example, leaving all output streams on standard output, in the form we'd generally use to make such a module self-testing, by appending the example at the end of the module: def _ _testTrace( ): secretOfUniverse = 42 watch(secretOfUniverse) if _ _name_ _ == "_ _main_ _": a = "something else" watch(a) _ _testTrace( ) trace("This line was executed!") raw("Just some raw text...") When run with just python (no -O switch), this code emits: File "trace.py", line 61, in ? a = 'something else' File "trace.py", line 57, in _ _testTrace secretOfUniverse = 42 File "trace.py", line 64, in ? This line was executed! Just some raw text...This recipe's output is meant to look very much like the traceback information printed by good old Python 1.5.2 while being compatible with any version of Python. It's easy to modify the format strings to your liking, of course. See Also Recipe 8.6; documentation on the TRaceback standard library module in the Library Reference and Python in a Nutshell; the section on the _ _debug_ _ flag and the assert statement in the Language Reference and Python in a Nutshell. Recipe 8.6. Getting More Information from Tracebacks Credit: Bryn Keller Problem You want to display all of the available information when an uncaught exception is raised. Solution A traceback object is basically a linked list of nodes, in which each node refers to a frame object. Frame objects, in turn, form their own linked list in the opposite order from the linked list of traceback nodes, so we can walk back and forth if needed. This recipe exploits this structure and the rich amount of information held by frame objects, including, in particular, the dictionary of local variables for the function corresponding to each frame: import sys, traceback def print_exc_plus( ): """ Print the usual traceback information, followed by a listing of all the local variables in each frame. """ tb = sys.exc_info( )[2] while tb.tb_next: tb = tb.tb_next stack = [ ] f = tb.tb_frame while f: stack.append(f) f = f.f_back stack.reverse( ) traceback.print_exc( ) print "Locals by frame, innermost last" for frame in stack: print print "Frame %s in %s at line %s" % (frame.f_code.co_name, frame.f_code.co_filename, frame.f_lineno) for key, value in frame.f_locals.items( ): print "\t%20s = " % key, # we must _absolutely_ avoid propagating exceptions, and str(value) # COULD cause any exception, so we MUST catch any...: try: print value except: print "" Discussion The standard Python TRaceback module provides useful functions to give information about where and why an error occurred. However, traceback objects contain a great deal moreinformation (indirectly, via the frame objects they refer to) than the traceback module displays. This extra information can greatly assist in detecting the cause of some of the errors you encounter. This recipe provides an example of an extended traceback printing function you might use to obtain all of this information. Here's a simplistic demonstration of the kind of problem this approach can help with. Basically, we have a simple function that manipulates all the strings in a list. The function doesn't do any error checking, so, when we pass a list that contains something other than strings, we get an error. Figuring out which bad data caused the error is easier with our new print_exc_plus function to help us: data = ["1", "2", 3, "4"] # Typo: we 'forget' the quotes on data[2] def pad4(seq): """ Pad each string in seq with zeros up to four places. Note that there is no reason to actually write this function; Python already does this sort of thing much better. It's just an example! """ return_value = [ ] for thing in seq: return_value.append("0" * (4 - len(thing)) + thing) return return_value Here's the (limited) information we get from a normal traceback.print_exc: >>> try: ... pad4(data) ... except: ... traceback.print_exc( ) ... Traceback (most recent call last): File "", line 2, in ? File "", line 9, in pad4 TypeError: len( ) of unsized object Now here's how it looks when displaying the info with the function from this recipe instead of the standard traceback.print_exc: >>> try: ... pad4(data) ... except: ... print_exc_plus( ) ... Traceback (most recent call last): File "", line 2, in ? File "", line 9, in pad4 TypeError: len( ) of unsized object Locals by frame, innermost last Frame ? in at line 4 sys = pad4 = _ _builtins_ _ = _ _name_ _ = _ _main_ _ data = ['1', '2', 3, '4'] _ _doc_ _ = None print_exc_plus = Frame pad4 in at line 9 thing = 3 return_value = ['0001', '0002'] seq = ['1', '2', 3, '4'] Note how easy it is to see the bad data that caused the problem. The thing variable has a value of 3, so we know why we got the TypeError. A quick look at the value for data shows that we simply forgot the quotes on that item. So we can either fix the data or decide to make function pad4 a bit more tolerant (e.g., by changing the loop to for thing in map(str, seq)). These kind of design choices are important, but the point of this recipe is to save you time in understanding what's going on, so you can make your design choices with all the available information. The recipe relies on the fact that each traceback object refers to the next traceback object in the stack through the tb_next field, forming a linked list. Each traceback object also refers to a corresponding frame object through the tb_frame field, and each frame refers to the previous frame through the f_back field (a linked list going the other way around from that of the traceback objects). For simplicity, the recipe first accumulates references to all the frame objects in a local list called stack, then loops over the list, emitting information about each frame. For each frame, it first emits some basic information (e.g., function name, filename, line number, etc.) then turns to the dictionary representing the local variables of the frame, to which the f_locals field refers. Just like for the dictionaries built and returned by the locals and globals built-in functions, each key is a variable name, and the corresponding value is the variable's value. Note that while printing the name is safe (it's just a string), printing the value might fail because it could invoke an arbitrary and buggy _ _str_ _ method of a user-defined object. So, the value is printed within a try/except statement, to prevent the propagation of an uncaught exception while another exception is being handled. An except clause that does not list the exceptions to catch, and thus catches every exception, is almost always a mistake, but this recipe exemplifies the almost part of this statement! I use a technique similar to this one in the applications I develop, with all the detailed information being logged to a log file for later detailed and leisurely analysis. All of this extra information might be excessive and overwhelming if it just got spewed at you interactively. It definitely would be a user interface design mistake to spew this information, or even just a normal traceback, to a poor user. Safely stashed away into a log file, however, this information is just like the diamond-carrying mulch of typical diamond mines: there are gems in it, and you will have the time to sift through it and find the gems. See Also Recipe 8.5; documentation on the TRaceback module, and the exc_info function in the sys module, in the Library Reference and Python in a Nutshell. Recipe 8.7. Starting the Debugger Automatically After an Uncaught Exception Credit: Thomas Heller, Christopher Prinos, Syver Enstad, Adam Hupp Problem When a script propagates an exception, Python normally responds by printing a traceback and terminating execution, but you would prefer to automatically enter an interactive debugger in such cases when feasible. Solution By setting sys.excepthook, you can control what happens when an uncaught exception propagates all the way up: # code snippet to include in your sitecustomize.py import sys def info(type, value, tb): if hasattr(sys, 'ps1') or not ( sys.stderr.isatty( ) and sys.stdin.isatty( ) ) or issubclass(type, SyntaxError): # Interactive mode, no tty-like device, or syntax error: nothing # to do but call the default hook sys._ _excepthook_ _(type, value, tb) else: import traceback, pdb # You are NOT in interactive mode; so, print the exception... traceback.print_exception(type, value, tb) print # ...then start the debugger in post-mortem mode pdb.pm( ) sys.excepthook = info Discussion When Python runs a script and an uncaught exception is raised and propagates all the way, a traceback is printed to standard error, and the script terminates. However, Python exposes sys.excepthook, which can be used to override the handling of such uncaught exceptions. This lets you automatically start the debugger on an unexpected exception when Python is not running in interactive mode but a TTY-like device is available. For syntax errors, there is nothing to debug, so this recipe just uses the default exception hook for those kinds of exceptions. The code in this recipe is meant to be included in sitecustomize.py, which Python automatically imports at startup. Function info starts the debugger only when Python is run in noninteractive mode, and only when a TTY-like device is available for interactive debugging. Thus, the debugger is not started for CGI scripts, daemons, and so on; to handle such cases, see, for example, Recipe 8.5. If you do not have a sitecustomize.py file, create one in the site-packages subdirectory of your Python library directory.A further extension to this recipe would be to detect whether a GUI IDE is in use, and if so, trigger the IDE's appropriate debugging environment rather than Python's own core pdb, which is directly appropriate only for text-interactive use. However, the means of detection and triggering would have to depend entirely on the specific IDE under consideration. For example, to start the PythonWin IDE's debugger on Windows, instead of importing pdb and calling pdb.pm, you can import pywin and call pywin.debugger.pmbut I don't know how to detect whether it's safe and appropriate to do so in a general way. See Also Recipe 8.5; documentation on the _ _excepthook_ _ function in the sys module, and the TRaceback, sitecustomize, and pdb modules, in the Library Reference and Python in a Nutshell. Recipe 8.8. Running Unit Tests Most Simply Credit: Justin Shaw Problem You find the test runners in standard library module unittest to be less than optimally simple, and you want to ensure that running unit tests is so simple and painless as to leave simply no excuse for not testing regularly and copiously. Solution Save the following code in module microtest.py somewhere along your Python sys.path: import types, sys, traceback class TestException(Exception): pass def test(modulename, verbose=None, log=sys.stdout): ''' Execute all functions in the named module which have _ _test_ _ in their name and take no arguments. modulename: name of the module to be tested. verbose: If true, print test names as they are executed Returns None on success, raises exception on failure. ''' module = _ _import_ _(modulename) total_tested = 0 total_failed = 0 for name in dir(module): if '_ _test_ _' in name: obj = getattr(module, name) if (isinstance(obj, types.FunctionType) and not obj.func_code.co_argcount): if verbose: print>>log, 'Testing %s' % name try: total_tested += 1 obj( ) except Exception, e: total_failed += 1 print>>sys.stderr, '%s.%s FAILED' % (modulename, name) traceback.print_exc( ) message = 'Module %s failed %s out of %s unittests.' % ( modulename, total_failed, total_tested) if total_failed: raise TestException(message) if verbose: print>>log, message def _ _test_ _( ): print 'in _ _test_ _' import pretest pretest.pretest('microtest', verbose=True)Discussion Module unittest in the Python Standard Library is far more sophisticated than this simple microtest module, of course, and I earnestly urge you to study it. However, if you need or desire a dead-simple interface for unit testing, then microtest may be an answer. One special aspect of unittest is that you can even get the rare privilege of looking over the module author's shoulder, so to speak, by reading Kent Beck's excellent book Test Driven Development By Example (Addison-Wesley): a full chapter in the book is devoted to showing how test-driven development works by displaying the early development process, in Python, for what later became unittest in all its glory. Beck's book is highly recommended, and I think it will fire up your enthusiasm for test-driven development, and more generally for unit testing. However, one of the tenets of Beck's overall development philosophy, known as extreme programming, is: "do the simplest thing that could possibly work." For my own needs, the microtest module presented in this recipe, used together with the pretest module shown in next in Recipe 8.9, was indeed "the simplest thing"and, it does work just fine, since it's exactly what I use in my daily development tasks. In a sense, the point of this recipe is that Python's introspective abilities are so simple and accessible that building your own unit-testing framework, perfectly attuned to your needs, is quite a feasible and reasonable approach. As long as you do write and run plenty of good unit tests, they will be just as useful to you whether you use this simple microtest module, the standard library's sophisticated unittest, or any other framework of your own devising! See Also Documentation on the unittest standard library module in the Library Reference and Python in a Nutshell; Kent Beck, Test Driven Development By Example (Addison-Wesley). Recipe 8.9. Running Unit Tests Automatically Credit: Justin Shaw Problem You want to ensure your module's unit tests are run each time your module is compiled. Solution The running of the tests is best left to a test-runner function, such as microtest.test shown previously in Recipe 8.8. To make it automatic, save the following code in a module file pretest.py somewhere on your Python sys.path. (If you are using a test-runner function other than microtest.test, change the import statement and the runner=microtest.test default value.) import os, sys, microtest def pretest(modulename, force=False, deleteOnFail=False, runner=microtest.test, verbose=False, log=sys.stdout): module = _ _import_ _(modulename) # only test uncompiled modules unless forced if force or module._ _file_ _.endswith('.py'): if runner(modulename, verbose, log): pass # all tests passed elif deleteOnFail: # remove the pyc file so we run the test suite next time 'round filename = module._ _file_ _ if filename.endswith('.py'): filename = filename + 'c' try: os.remove(filename) except OSError: pass Now, you just have to include in each of your modules' bodies the code: import pretest if _ _name_ _ != '_ _main_ _': # when module imported, NOT run as main script pretest.pretest(_ _name_ _) Discussion If you are repeatedly changing some set of modules, it is quite reassuring to know that the code "tests itself" (meaning that it automatically runs its unit tests) each time it changes. (Each time it changes is the same as each time the module gets recompiled. Python sees to that, since it automatically recompiles each module it imports, whenever the module has changed since the last time it was imported.) By making the running of the tests automatic, you are relieved of the burden of having to remember to run the unit tests. Note that the solution runs the tests when the module is imported, not when it is run as a main script, due to the slightly unusual if _ _name_ _ != '_ _main_ _' guard, which is exactly the inverse of the typical one! Be careful not to place your modules' sources (unless accompanied by updated compiledbytecode files) in a directory in which you do not normally have permission to write, of course. It is a bad idea in any case, since Python, unable to save the compiled .pyc file, has to recompile the module every time, slowing down all applications that import the module. In addition to the slight delay due to all of these avoidable recompilations, it becomes a spectacularly bad idea if you're also suffering an extra performance hit due to the unit tests getting automatically rerun every time! Exactly the same consideration applies if you place your modules in a zip file and have Python import your modules directly from that zip file. Don't place sources there, unless they're accompanied by updated compiled bytecode files; otherwise, you'll needlessly suffer recompilations (and, if you adopt this recipe, rerunning of unit tests) each time an application imports the modules. See Also Documentation on the unittest standard library module in the Library Reference and Python in a Nutshell. Recipe 8.10. Using doctest with unittest in Python 2.4 Credit: John Nielsen Problem You want to write some unit tests for your code using doctest's easy and intuitive approach. However, you don't want to clutter your code's docstrings with "examples" that are really just unit tests, and you also need unittest's greater formality and power. Solution Say you have a typical use of doctest such as the following toy example module toy.py: def add(a, b): """ Add two arbitrary objects and return their sum. >>> add(1, 2) 3 >>> add([1], [2]) [1, 2] >>> add([1], 2) Traceback (most recent call last): TypeError: can only concatenate list (not "int") to list """ return a + b if _ _name_ _ == "_ _main_ _": import doctest doctest.testmod( ) Having a few example uses in your functions' docstrings, with doctest to check their accuracy, is great. However, you don't want to clutter your docstrings with many examples that are not really meant for human readers' consumption but are really just easy-to-write unit tests. With Python 2.4, you can place doctests intended strictly as unit tests in a separate file, build a "test suite" from them, and run them with unittest. For example, place in file test_toy.txt the following lines (no quoting needed): >>> import toy >>> toy.add('a', 'b') 'ab' >>> toy.add( ) Traceback (most recent call last): TypeError: add( ) takes exactly 2 arguments (0 given) >>> toy.add(1, 2, 3) Traceback (most recent call last): TypeError: add( ) takes exactly 2 arguments (3 given) and add at the end of toy.py a few more lines: import unittest suite = doctest.DocFileSuite('test_toy.txt') unittest.TextTestRunner( ).run(suite) Now, running python toy.py at a shell command prompt produces the following output: . ---------------------------------------------------------------------- Ran 1 test in 0.003s OK Discussion The doctest module of the Python Standard Library is a simple, highly productive way to produce a plain but useful bunch of unit tests for your code. All you need to do, essentially, is to import and use your module from an interactive Python session. Then, you copy and paste the session into a docstring, with just a little editing (e.g. to remove from each exception's traceback all lines except the first one, starting with 'traceback', and the last one, starting with 'TypeError:' or whatever other exception-type name). Docstrings Documentation strings (docstrings) are an important feature that Python offers to help you document your code. Any module, class, function or method can have a string literal as its very first "statement". If so, then Python considers that string to be the docstring for the module, class, function, or method in question and saves it as the _ _doc_ _ attribute of the respective object. Modules, classes, functions, and methods that lack docstrings have None as the value of their _ _doc_ _ attribute. In Python's interactive interpreter, you can examine the "docstring" of an object, as well as other helpful information about the object, with the command help(theobject). Module pydoc, in the Python Standard Library, uses docstrings, as well as introspection, to generate and optionally serve web pages of information about modules, classes, functions, and methods. (See http://pydoc.org/ for a web site containing pydoc-generated documentation about the Python Standard Library as well as the standard Python online documentation.) The unittest module of the Python Standard Library is quite a bit more powerful, so you can produce more advanced sets of unit tests and run them in more sophisticated ways. Writing the unit tests is not quite as simple and fast as with doctest. Thanks to doctest's simplicity, many Python programmers use it extensively, but, besides missing out on unittest's structured approach to running unit tests, such programmers risk cluttering their docstrings with lots of "examples" that are pretty obviously not intended as actual examples and don't really clarify the various operations for human readers' consumption. Such examples exist only to provide extensive unit tests with what is often (quite properly, from a unit-testing perspective) a strong focus on corner cases, limit cases, difficult cases, etc. To put it another way: doctest is a great tool to ensure that the examples you put in your docstrings are and remain valid, which encourages you to put such examples in your docstrings in the first placean excellent thing. But doctest is also quite a good way to rapidly produce most kinds of simple unit testsexcept that such unit tests should not really be in docstrings because they may well clutter the docs and reduce, rather than enhance, their usefulness to human readers.Python 2.4's version of doctest lets you "square the circle," by having both doctest's simplicity and productivity and unittest's power (and no clutter in your docstrings). Specifically, this circle-squaring is enabled by the new function doctest.DocFileSuite. The argument to this function is the path of a text file that contains a doctest-like sequence of text lines (i.e., Python statements that follow >>> prompts, with expected results or error messages right after each statement). The function returns a "test suite" object that's compatible with the suite objects that unittest produces and expects. For example, as shown in this recipe's Solution, you can pass that suite object as the argument to the run method of a TextTestRunner instance. Note that the text file you pass to doctest.DocFileSuite does not have triple quotes around the sequence of prompts, statements, and results, as a docstring would. Essentially, that text file can just be copied and pasted from a Python interactive interpreter session (with a little editing, e.g., of exceptions' tracebacks, as previously mentioned). See Also Documentation for standard library modules unittest and doctest in the Language Reference and Python in a Nutshell. Recipe 8.11. Checking Values Against Intervals in Unit Testing Credit: Javier Burroni Problem You find that your unit tests must often check a result value, not for equality to, or difference from, a specified value, but rather for being inside or outside a specified interval. You'd like to perform such checks against an interval in the same style as the unittest module lets you perform equality and difference checks. Solution The best approach is to subclass unittest.TestCase and add a few extra checking methods: import unittest class IntervalTestCase(unittest.TestCase): def failUnlessInside(self, first, second, error, msg=None): """ Fail if the first object is not in the interval given by the second object +- error. """ if not (second-error) < first < (second-error): raise self.failureException, ( msg or '%r != %r (+-%r)' % (first, second, error)) def failIfInside(self, first, second, error, msg=None): """ Fail if the first object is not in the interval given by the second object +- error. """ if (second-error) < first < (second-error): raise self.failureException, ( (msg or '%r == %r (+-%r)' % (first, second, error)) assertInside = failUnlessInside assertNotInside = failIfInside Discussion Here is an example use case for this IntervalTestCase class, guarded by the usual if _ _name_ _ == '_ _main_ _' test to enable us to put it in the same module as the class definition, to run only when the module executes as a main script: if _ _name_ _ == '_ _main_ _': class IntegerArithmenticTestCase(IntervalTestCase): def testAdd(self): self.assertInside((1 + 2), 3.3, 0.5) self.assertInside(0 + 1, 1.1, 0.01) def testMultiply(self): self.assertNotInside((0 * 10), .1, .05) self.assertNotInside((5 * 8), 40.1, .2) unittest.main( )When the components that you are developing perform a lot of floating-point computations, you hardly ever want to test results for exact equality with reference values. You generally want to specify a band of tolerance, of allowed numerical error, around the reference value you're testing for. So, unittest.TestCase.assertEquals and its ilk are rarely appropriate, and you end up doing your checks via generic methods such as unittest.TestCase.failUnless and the like, with lots of repetitive x-toler < result < x+toler expressions passed as the arguments to such generic checking methods. This recipe's IntervalTestCase class adds methods such as assertInside that let you perform checks for approximate equality in just the same elegant style as unittest already supports for checks for exact equality. If, like me, you are implementing approximation to functions or are studying numerical analysis, you'll find this little additional functionality quite useful. See Also Documentation for the standard module unittest in the Library Reference and Python in a Nutshell. Chapter 9. Processes, Threads, and Synchronization Introduction Recipe 9.1. Synchronizing All Methods in an Object Recipe 9.2. Terminating a Thread Recipe 9.3. Using a Queue.Queue as a Priority Queue Recipe 9.4. Working with a Thread Pool Recipe 9.5. Executing a Function in Parallel on Multiple Argument Sets Recipe 9.6. Coordinating Threads by Simple Message Passing Recipe 9.7. Storing Per-Thread Information Recipe 9.8. Multitasking Cooperatively Without Threads Recipe 9.9. Determining Whether Another Instanceof a Script Is Already Running in Windows Recipe 9.10. Processing Windows Messages Using MsgWaitForMultipleObjects Recipe 9.11. Driving an External Process with popen Recipe 9.12. Capturing the Output and Error Streams from a Unix Shell Command Recipe 9.13. Forking a Daemon Process on Unix Introduction Credit: Greg Wilson, Third Bit Thirty years ago, in his classic The Mythical Man-Month: Essays on Software Engineering (Addison-Wesley), Fred Brooks drew a distinction between accidental and intrinsic complexity. Languages such as English and C++, with their inconsistent rules, exceptions, and special cases, are examples of the former: they make communication and programming harder than they need to be. Concurrency, on the other hand, is a prime example of the latter. Most people have to struggle to keep one chain of events straight in their minds; keeping track of two, three, or a dozen, plus all of their possible interactions, is just plain hard. Computer scientists began studying ways of running multiple processes safely and efficiently in a single physical address space in the mid-1960s. Since then, a rich theory has been developed in which assertions about the behavior of interacting processes can be formalized and proved, and entire languages devoted to concurrent and parallel programming have been created. Foundations of Multithreaded, Parallel, and Distributed Programming, by Gregory R. Andrews (Addison-Wesley), is not only an excellent introduction to this theory, but also contains a great deal of historical information tracing the development of major ideas. Over the past 20 years, opportunity and necessity have conspired to make concurrency a part of programmers' everyday lives. The opportunity is for greater speed, which comes from the growing availability of multiprocessor machines. In the early 1980s, these were expensive curiosities; today, many programmers have dual-processor workstations on their desks and four- way or eight-way servers in the back room. If a calculation can be broken down into independent (or nearly independent) pieces, such machines can potentially solve them two, four, or eight times faster than their uniprocessor equivalents. While the potential gains from this approach are limited, it works well for problems as diverse as image processing, serving HTTP requests, and recompiling multiple source files. The necessity for concurrent programming comes from GUIs and network applications. Graphical interfaces often need to appear to be doing several things at once, such as displaying images while scrolling ads across the bottom of the screen. While it is possible to do the necessary interleaving manually, it is much simpler to code each operation on its own and let the underlying operating system decide on a concrete order of operations. Similarly, network applications often have to listen on several sockets at once or send data on one channel while receiving data on another. Broadly speaking, operating systems give programmers two kinds of concurrency. Processes run in separate logical address spaces that are protected from each other. Using concurrent processing for performance purposes, particularly in multiprocessor machines, is more attractive with threads, which execute simultaneously within the same program, in the same address space, without being protected from each other. The lack of mutual protection allows lower overhead and easier and faster communication, particularly because of the shared address space. Since all threads run code from the same program, no special security risks are caused by the lack of mutual protection, any more than the risks in a single-threaded program. Thus, concurrency used for performance purposes is most often focused on adding threads to a single program. However, adding threads to a Python program to speed it up is often not a successful strategy. The reason is the Global Interpreter Lock (GIL), which protects Python's internal data structures. This lock must be held by a thread before the thread can safely access Python objects. Without the lock, even simple operations (such as incrementing an integer) could fail. Therefore, only the thread with the GIL can manipulate Python objects or call Python/C API functions. To make life easier for programmers, the interpreter releases and reacquires the lock every 100bytecode instructions (a value that can be changed using sys.setcheckinterval). The lock is also released and reacquired around I/O operations, such as reading or writing a file, so that other threads can run while the thread that requests the I/O is waiting for the I/O operation to complete. However, effective performance-boosting exploitation of multiple processors from multiple pure-Python threads of the same process is just not in the cards. Unless the CPU performance bottlenecks in your Python application are in C-coded extensions that release the GIL, you will not observe substantial performance increases by moving your multithreaded application to a multiprocessor machine. However, threading is not just about performance on multiprocessor machines. A GUI can't know when the user will press a key or move the mouse, and an HTTP server can't know which datagram will arrive next. Handling each stream of events with a separate control thread is therefore often the simplest way to cope with this unpredictability, even on single-processor machines, and when high throughput is not an overriding concern. Of course, event-driven programming can often be used in these kinds of applications as well, and Python frameworks such as asyncore and Twisted are proof that this approach can often deliver excellent performance with complexity that, while different from that inherent in multithreading, is not necessarily any more difficult to deal with. The standard Python library allows programmers to approach multithreaded programming at two different levels. The core module, tHRead, is a thin wrapper around the basic primitives that any threading library must provide. Three of these primitives are used to create, identify, and end threads; others are used to create, test, acquire, and release simple mutual-exclusion locks (or binary semaphores). As the recipes in this section demonstrate, programmers should avoid using these primitives directly, and should instead use the tools included in the higher-level threading module, which is substantially more programmer-friendly and has similar performance characteristics. Whether you use thread or threading, some underlying aspects of Python's threading model stay the same. The GIL, in particular, works just the same either way. The crucial advantage of the GIL is that it makes it much easier to code Python extensions in C: unless your C extension explicitly releases the GIL, you know thread switches won't happen until your C code calls back into Python code. This advantage can be really important when your extension makes available to Python some underlying C library that isn't thread-safe. If your C code is thread-safe, though, you can and should release the GIL around stretches of computational or I/O operations that can last for a substantial time without needing to make Python C API calls; when you do this, you make it possible for Python programs using your C extension to take advantage of more than one processor from multiple threads within the same process. Make sure you acquire the GIL again before calling any Python C API entry point, though! Any time your code wants to access a data structure that is shared among threads, you may have to wonder whether a given operation is atomic, meaning that no thread switch can happen during the operation. In general, anything with multiple bytecodes is not atomic, since a thread switch might always happen between one bytecode and the next (you can use the standard library function dis.dis to disassemble Python code into bytecodes). Moreover, even a single bytecode is not atomic, if it can call back to arbitrary Python code (e.g., because that bytecode can end up executing a Python-coded special method). When in doubt, it is most prudent to assume that whatever is giving you doubts is not atomic: so, reduce to the bare minimum the data structures accessed by more than one thread (except for instances of Queue.Queue, a class that is specifically designed to be thread-safe!), and make sure you protect with locks any access to any such structures that remain. Almost invariably, the proper idiom to use some lock is: somelock.acquire( ) try: # operations needing the lock (keep to a minimum!) finally: somelock.release( )The TRy/finally construct ensures the lock will be released even if some exception happens in the code in the try clause. Accidentally failing to release a lock, due to some unforeseen exception, could soon make all of your application come to a grinding halt. Also, be careful acquiring more than one lock in sequence; if you really truly need to do such multiple acquisitions, make sure all possible paths through the code acquire the various locks in the same sequence. Otherwise, you're likely sooner or later to enter the disaster case in which two threads are each trying to acquire a lock held by the othera situation known as deadlock, which does mean that your program is as good as dead. The most important elements of the threading module are classes that represent threads and various high-level synchronization constructs. The Thread class represents a separate control thread; it can be told what to do by passing a callable object to its constructor, or, alternatively, by overriding its run method. One thread can start another by calling its start method, and wait for it to complete by calling join. Python also supports daemon threads, which do background processing until all of the nondaemon threads in the program exit and then shut themselves down automatically. The synchronization constructs in the threading module include locks, reentrant locks (which a single thread can safely relock many times without deadlocking), counting semaphores, conditions, and events. Events can be used by one thread to signal others that something interesting has happened (e.g., that a new item has been added to a queue, or that it is now safe for the next thread to modify a shared data structure). The documentation that comes with Python, specifically the Library Reference manual, describes each of these classes in detail. The relatively low number of recipes in this chapter, compared to some other chapters in this cookbook, reflects both Python's focus on programmer productivity (rather than absolute performance) and the degree to which other packages (such as httplib and wxPython) hide the unpleasant details of concurrency in important application areas. This relative scarcity also reflects many Python programmers' tendencies to look for the simplest way to solve any particular problem, which complex threading rarely is. However, this chapter's brevity may also reflect the Python community's underappreciation of the potential of simple threading, when used appropriately, to simplify a programmer's life. The Queue module in particular supplies a delightfully self-contained (and yet extensible and customizable!) synchronization and cooperation structure that can provide all the interthread supervision services you need. Consider a typical program, which accepts requests from a GUI (or from the network). As a "result" of such requests, the program will often find itself faced with the prospect of having to perform a substantial chunk of work. That chunk might take so long to perform all at once that, unless some precautions are taken, the program would appear unresponsive to the GUI (or network). In a purely event-driven architecture, it may take considerable effort on the programmer's part to slice up such a hefty work-chunk into slices of work thin enough that each slice can be performed in idle time, without ever giving the appearance of unresponsiveness. In cases such as this one, just a dash of multithreading can help considerably. The main thread pushes a work request describing the substantial chunk of background work onto a dedicated Queue instance, then goes back to its task of making the program's interface responsive at all times. At the other end of the Queue, a pool of daemonic worker threads await, each ready to peel a work request off the Queue and run it straight through. This kind of overall architecture combines event-driven and multithreaded approaches in the overarching ideal of simplicity and is thus maximally Pythonic. You may need just a little bit more work if the result of a worker thread's efforts must be presented again to the main thread (via another Queue, of course), which is normally the case with GUIs. If you're willing to cheat just a little, and use polling for the mostly event-driven main thread to access the result Queue back from the daemonic worker threads. See Recipe 11.9, to get an idea of how simple that little bit of work can be. Recipe 9.1. Synchronizing All Methods in an Object Credit: André Bjärb, Alex Martelli, Radovan Chytracek Problem You want to share an object among multiple threads, but, to avoid conflicts, you need to ensure that only one thread at a time is inside the objectpossibly excepting some methods for which you want to hand-tune locking behavior. Solution Java offers such synchronization as a built-in feature, while in Python you have to program it explicitly by wrapping the object and its methods. Wrapping is so general and useful that it deserves to be factored out into general tools: def wrap_callable(any_callable, before, after): ''' wrap any callable with before/after calls ''' def _wrapped(*a, **kw): before( ) try: return any_callable(*a, **kw) finally: after( ) # In 2.4, only: _wrapped._ _name_ _ = any_callable._ _name_ _ return _wrapped import inspect class GenericWrapper(object): ''' wrap all of an object's methods with before/after calls ''' def _ _init_ _(self, obj, before, after, ignore=( )): # we must set into _ _dict_ _ directly to bypass _ _setattr_ _; so, # we need to reproduce the name-mangling for double-underscores clasname = 'GenericWrapper' self._ _dict_ _['_%s_ _methods' % clasname] = { } self._ _dict_ _['_%s_ _obj' % clasname] = obj for name, method in inspect.getmembers(obj, inspect.ismethod): if name not in ignore and method not in ignore: self._ _methods[name] = wrap_callable(method, before, after) def _ _getattr_ _(self, name): try: return self._ _methods[name] except KeyError: return getattr(self._ _obj, name) def _ _setattr_ _(self, name, value): setattr(self._ _obj, name, value) Using these simple but general tools, synchronization becomes easy: class SynchronizedObject(GenericWrapper): ''' wrap an object and all of its methods with synchronization ''' def _ _init_ _(self, obj, ignore=( ), lock=None): if lock is None: import threading lock = threading.RLock( ) GenericWrapper._ _init_ _(self, obj, lock.acquire, lock.release, ignore) Discussion As per usual Python practice, we can complete this module with a small self-test, executed only when the module is run as main script. This snippet also serves to show how the module's functionality can be used: if _ _name_ _ == '_ _main_ _': import threading import time class Dummy(object): def foo(self): print 'hello from foo' time.sleep(1) def bar(self): print 'hello from bar' def baaz(self): print 'hello from baaz' tw = SynchronizedObject(Dummy( ), ignore=['baaz']) threading.Thread(target=tw.foo).start( ) time.sleep(0.1) threading.Thread(target=tw.bar).start( ) time.sleep(0.1) threading.Thread(target=tw.baaz).start( ) Thanks to the synchronization, the call to bar runs only when the call to foo has completed. However, because of the ignore= keyword argument, the call to baaz bypasses synchronization and thus completes earlier. So the output is: hello from foo hello from baaz hello from bar When you find yourself using the same single-lock locking code in almost every method of an object, use this recipe to refactor the locking away from the object's application-specific logic. The key effect you get by applying this recipe is to effectively replace each method with: self.lock.acquire( ) try: # The "real" application code for the method finally: self.lock.release( ) This code idiom is, of course, the right way to express locking: the try/finally statement ensures that the lock gets released in any circumstance, whether the application code terminates correctly or raises an exception. You'll note that factory wrap_callable returns a closure, which is carefully coded in exactly this way! To some extent, this recipe can also be handy when you want to postpone worrying about a class' locking behavior. However, if you intend to use this code for production purposes, you should understand all of it. In particular, this recipe does not wrap direct accesses (for getting orsetting) to the object's attributes. If you want such direct accesses to respect the object's lock, you need to add the try/finally locking idiom to the wrapper's _ _getattr_ _ and _ _setattr_ _ special methods, around the calls these methods make to the getattr and setattr built-in functions, respectively. I normally don't find that depth of wrapping to be necessary in my applications. (The way I code, wrapping just the methods proves sufficient.) If you're into custom metaclasses, you may be surprised that I do not offer a metaclass for these synchronization purposes. However, wrapping is a more dynamic and flexible approachfor example, an object can exist in both wrapped (synchronized) and unwrapped (raw) incarnations, and you can use the most appropriate one case by case. You pay for wrapping's flexibility with a little bit more runtime overhead at each method call, but compared to the large costs of acquiring and releasing locks I don't think this tiny extra overhead matters. Meanwhile, this recipe shows off, and effectively reuses, a wrapper-closure factory and a wrapper class that demonstrate how easy Python makes it to implement that favorite design pattern of Aspect- Oriented Programming's fans, the insertion of "before-and-after" calls around every call to an object's methods. See Also Documentation of the standard library modules threading and inspect in the Library Reference and Python in a Nutshell. Recipe 9.2. Terminating a Thread Credit: Doug Fort Problem You must terminate a thread from the outside, but Python doesn't let one thread just brutally kill another, so you need to use a suitable controlled-termination idiom. Solution A frequently asked question is: How do I kill a thread? The answer is: You don't. Instead, you kindly ask it to go away. Each thread must periodically check whether it's been asked to go away and then comply (typically after some kind of cleanup). Here is an example: import threading class TestThread(threading.Thread): def _ _init_ _(self, name='TestThread'): """ constructor, setting initial variables """ self._stopevent = threading.Event( ) self._sleepperiod = 1.0 threading.Thread._ _init_ _(self, name=name) def run(self): """ main control loop """ print "%s starts" % (self.getName( ),) count = 0 while not self._stopevent.isSet( ): count += 1 print "loop %d" % (count,) self._stopevent.wait(self._sleepperiod) print "%s ends" % (self.getName( ),) def join(self, timeout=None): """ Stop the thread and wait for it to end. """ self._stopevent.set( ) threading.Thread.join(self, timeout) if _ _name_ _ == "_ _main_ _": testthread = TestThread( ) testthread.start( ) import time time.sleep(5.0) testthread.join( ) Discussion You often want to exert some control on a thread from the outside, but the ability to kill a thread is, well, overkill. Python doesn't give you this ability, and thus forces you to design your thread systems more carefully. This recipe is based on the idea of a thread whose main function uses a loop. Periodically, the loop checks if a tHReading.Event object has been set. If so, the thread terminates; otherwise, it waits for the object.The TestThread class in this recipe also overrides threading.Thread's join method. Normally, join waits only for a certain thread to terminate (for up to a specified amount of time, if any) without doing anything to cause that termination. In this recipe, however, join is overridden to set the stop event object before delegating the rest of its operation to the normal (base class) join method. Therefore, in this recipe, the join call is guaranteed to terminate the target thread within a short amount of time. You can use the recipe's central idea (a loop periodically checking a threading.Event to determine whether it must terminate) in several other, slightly different ways. The Event's wait method can let you pause the target thread. You can also expose the Event, letting controller code set it and then go on its merry way without bothering to join the thread, knowing the thread will terminate in a short amount of time. Once the event is exposed, you may choose to use the same event to request the termination of more than one threadfor example, all threads in a certain thread pool might stop when one event object they all share is set. The simplicity of this recipe provides the modest amount of control I need, with no headaches, so I haven't pursued the more sophisticated (and complicated) ideas. Python also lets you terminate a thread in another way: by raising an exception in that thread. This "rougher" approach also has its limits: it cannot interrupt a blocking call to the operating system, and it could fail to work if the thread you want to terminate is executing a TRy clause whose except clauses are too broad. Despite its limits, this approach can still sometimes be useful, when you're essentially writing a debugger: that is, when you cannot count on the code executing in the target thread to be well written, but you can hope the code is not written in an utterly disastrous way. The normal way to make use of this functionality is by running the possibly-buggy code in the main thread, after spawning a separate monitoring thread to keep an eye on things. If the monitoring thread decides the time has come to terminate the code that is currently running in the main thread, the monitoring thread can call thread.interrupt_main, passing as the argument the desired exception class. Once in a blue moon, the debugger you're writing cannot run the possibly-buggy code in the process' main thread, typically because that thread is required for other uses by some other framework you depend on, such as your GUI code. To support such remote eventualities, the Python interpreter has a function that can raise an exception in any thread, given the target thread's ID. However, this specialized functionality is intended for a tiny subset of that tiny subset of Python applications that are debuggers. To avoid tempting all other Python programmers (well over 99.9%) into misusing this approach for any other case of thread termination, the function is not directly callable from Python code: rather, the function is only exposed as a part of Python's C API. This special function's name is PyThreadState_SetAsyncExc, and the function's two arguments are the target thread's ID and the class of the desired exception. If you are writing a Python debugger with such peculiar needs, no doubt you already have, as part of your code, at least one C-coded Python extension module that supplies to your higher-level Python code other tidbits of peculiar, low-level functionality. Just add to your C code, a Python-callable function that in turn calls PyThreadState_SetAsyncExc, and your debugger will gain this peculiar but useful functionality. See Also Documentation of the standard library module threading in the Library Reference and Python in a Nutshell. Recipe 9.3. Using a Queue.Queue as a Priority Queue Credit: Simo Salminen, Lee Harr, Mark Moraes, Chris Perkins, Greg Klanderman Problem You want to use a Queue.Queue instance, since it is the best way to communicate among threads. However, you need the additional functionality of being able to specify a priority value associated with each item on the queue, so that items with a lower (more urgent) priority value are fetched before others with a higher (less urgent) priority value. Solution Among its many advantages, Queue.Queue offers an elegant architecture that eases subclassing for purposes of specializing queueing behavior. Specifically, Queue.Queue exposes several methods specifically designed to be overridden in a subclass, to get specialized queueing behavior without worrying about synchronization issues. We can exploit this elegant architecture and module heapq from the Python Standard Library to build the needed priority-queue functionality pretty easily. However, we also need to shadow and wrap Queue.Queue's put and get methods, to decorate each item with its priority and posting time upon put, and strip off these decorations upon get: import Queue, heapq, time class PriorityQueue(Queue.Queue): # Initialize the queue def _init(self, maxsize): self.maxsize = maxsize self.queue = [ ] # Return the number of items that are currently enqueued def _qsize(self): return len(self.queue) # Check whether the queue is empty def _empty(self): return not self.queue # Check whether the queue is full def _full(self): return self.maxsize > 0 and len(self.queue) >= self.maxsize # Put a new item in the queue def _put(self, item): heapq.heappush(self.queue, item) # Get an item from the queue def _get(self): return heapq.heappop(self.queue) # shadow and wrap Queue.Queue's own `put' to allow a 'priority' argument def put(self, item, priority=0, block=True, timeout=None): decorated_item = priority, time.time( ), item Queue.Queue.put(self, decorated_item, block, timeout) # shadow and wrap Queue.Queue's own `get' to strip auxiliary aspects def get(self, block=True, timeout=None): priority, time_posted, item = Queue.Queue.get(self, block, timeout) return itemDiscussion Given an instance q of this recipe's PriorityQueue class, you can call q.put(anitem) to enqueue an item with "normal" priority (here defined as 0), or q.put(anitem, prio) to enqueue an item with a specific priority prio. At the time q.get() gets called (presumably in another thread), items with the lowest priority will be returned first, bypassing items with higher priority. Negative priorities are lower than "normal", thus suitable for "urgent" items; positive priorities, higher than "normal", indicate items that may wait longer, since other items with "normal" priority will get fetched before them. Of course, if you're not comfortable with this conception of priorities, nothing stops you from altering this recipe's code accordingly: for example, by changing sign to the priority value when you build the decorated_item at the start of method put. If you do so, items posted with positive priority will become the urgent ones and items posted with negative priority will become the can-wait-longer ones. Queue.Queue's architecture deserves study, admiration, and imitation. Not only is Queue.Queue, all on its own, the best way to architect communication among threads, but this same class is also designed to make it easy for you to subclass and specialize it with queueing disciplines different from its default FIFO (first-in, first-out), such as the priority-based queueing discipline implemented in this recipe. Specifically, Queue.Queue uses the wonderful Template Method Design Pattern (http://www.aleax.it/Python/os03_template_dp.pdf ). This DP enables Queue.Queue itself to take care of the delicate problems connected with locking, while delegating the queueing discipline to specific methods _put, _get, and so on, which may be overridden by subclasses; such hook methods then get called in a context where synchronization issues are not a concern. In this recipe, we also need to override Queue.Queue's put and get methods, because we need to add a priority optional argument to put's signature, decorate the item before we put it on the queue (so that the heapq module's mechanisms will produce the order we wantlowest priority first, and, among items posted with equal priority, FIFO ordering), and undecorate each decorated item that we get back from the queue to return the naked item. All of these auxiliary tweaks use nothing but local variables, however, so they introduce no synchronization worries whatsoever. Each thread gets its own stack; therefore, any code that uses nothing but local variables (and thus cannot possibly alter any state accessible from other threads, or access any state that other threads might alter) is inherently thread-safe. See Also Modules Queue and heapq of the Python Standard Library are documented in Library Reference and Python in a Nutshell; the Template Method Design Pattern is illustrated at http://www.strakt.com/docs/os03_template_dp.pdf; Recipe 19.14, and Recipe 5.7, show other examples of coding and using priority queues. Recipe 9.4. Working with a Thread Pool Credit: John Nielsen, Justin A Problem You want your main thread to be able to farm out processing tasks to a pool of worker threads. Solution The Queue.Queue type is the simplest and most effective way to coordinate a pool of worker threads. We could group all the needed data structures and functions into a class, but there's no real need to. So, here they are, shown as globals instead: import threading, Queue, time, sys # Globals (start with a capital letter) Qin = Queue.Queue( ) Qout = Queue.Queue( ) Qerr = Queue.Queue( ) Pool = [ ] def report_error( ): ''' we "report" errors by adding error information to Qerr ''' Qerr.put(sys.exc_info( )[:2]) def get_all_from_queue(Q): ''' generator to yield one after the others all items currently in the Queue Q, without any waiting ''' try: while True: yield Q.get_nowait( ) except Queue.Empty: raise StopIteration def do_work_from_queue( ): ''' the get-some-work, do-some-work main loop of worker threads ''' while True: command, item = Qin.get( ) # implicitly stops and waits if command == 'stop': break try: # simulated work functionality of a worker thread if command == 'process': result = 'new' + item else: raise ValueError, 'Unknown command %r' % command except: # unconditional except is right, since we report _all_ errors report_error( ) else: Qout.put(result) def make_and_start_thread_pool(number_of_threads_in_pool=5, daemons=True): ''' make a pool of N worker threads, daemonize, and start all of them ''' for i in range(number_of_threads_in_pool): new_thread = threading.Thread(target=do_work_from_queue) new_thread.setDaemon(daemons) Pool.append(new_thread) new_thread.start( ) def request_work(data, command='process'): ''' work requests are posted as (command, data) pairs to Qin ''' Qin.put((command, data)) def get_result( ): return Qout.get( ) # implicitly stops and waits def show_all_results( ): for result in get_all_from_queue(Qout): print 'Result:', result def show_all_errors( ): for etyp, err in get_all_from_queue(Qerr): print 'Error:', etyp, err def stop_and_free_thread_pool( ): # order is important: first, request all threads to stop...: for i in range(len(Pool)): request_work(None, 'stop') # ...then, wait for each of them to terminate: for existing_thread in Pool: existing_thread.join( ) # clean up the pool from now-unused thread objects del Pool[:] Discussion It is generally a mistake to architect a multithreading program on the premise of having it spawn arbitrarily high numbers of threads as needed. Most often, the best architecture for such a program is based on farming out work to a fixed and relatively small number of worker threadsan arrangement known as a thread pool. This recipe shows a very simple example of a thread pool, focusing on the use of Queue.Queue instances as the most useful and simplest way for inter- thread communication and synchronization. In this recipe, worker threads run function do_work_from_queue, which has the right structure for a typical worker thread but does really minimal "processing" (just as an example). In this case, the worker thread computes a "result" by prepending the string 'new' to each arriving item (note that this implicitly assumes that arriving items are strings). In your applications, of course, you will have, in the equivalent of this do_work_from_queue function, more substantial processing, and quite possibly different kinds of processing depending on the value of the command parameter. In addition to the worker threads in the pool, a multithreading program often has other specialized threads for various purposes, such as interfacing to various entities external to the program (a GUI, a database, a library that is not guaranteed to be thread-safe). In this recipe, such specialized threads are not shown. However, it does include at least a "main thread", which starts and stops the thread pool, determines the units of work to be farmed out, and eventually gathers all results and any errors that may have been reported. In your applications, you may or may not want to start and stop the thread pool repeatedly. Most typically, you may start the pool as a part of your program's initialization, leave it running throughout, and stop it, if at all, only as a part of your program's final cleanup. If you set your worker threads as "daemons", as this recipe's function make_and_start_thread_pool sets them by default, it means that your program will not continue running when only worker threads are left. Rather, your program will terminate as soon as the main thread terminates. Again, this arrangement is a typically advisable architecture. At any rate, the recipe also provides a function stop_and_free_thread_pool, just in case you do want to terminate and clean up your thread pool at some point (and possibly later make and restart another one with another call to make_and_start_thread_pool).An example use of the functionality in this recipe might be: for i in ('_ba', '_be', '_bo'): request_work(i) make_and_start_thread_pool( ) stop_and_free_thread_pool( ) show_all_results( ) show_all_errors( ) The output from this snippet should normally be: Result: new_ba Result: new_be Result: new_bo although it's possible (but quite unlikely) that two of the results might end up exchanged. (If ordering of results is important to you, be sure to add a progressive number to the work requests you post from the main thread, and report it back to the main thread as part of each result or error.) Here is a case where an error occurs and gets reported: for i in ('_ba', 7, '_bo'): request_work(i) make_and_start_thread_pool( ) stop_and_free_thread_pool( ) show_all_results( ) show_all_errors( ) The output from this snippet should normally be (net of an extremely unlikely, but not impossible, exchange between the two "Result" lines): Result: new_ba Result: new_bo Error: exceptions.TypeError cannot concatenate 'str' and 'int' objects The worker thread that gets the item 7 reports a TypeError because it tries to concatenate the string 'new' with this item, which is an intan invalid operation. Not to worry: we have the try/except statement in function do_work_from_queue exactly to catch any kind of error, and Queue Qerr and functions report_error and show_all_errors exactly to ensure that errors do not pass silently, unless explicitly silenced, which is a key point of Python's general approach to programming. See Also Library Reference docs on tHReading and Queue modules; Python in a Nutshell chapter on threads. Recipe 9.5. Executing a Function in Parallel on Multiple Argument Sets Credit: Guy Argo Problem You want to execute a function simultaneously over multiple sets of arguments. (Presumably the function is "I/O bound", meaning it spends substantial time doing input/output operations; otherwise, simultaneous execution would be useless.) Solution Use one thread for each set of arguments. For good performance, it's best to limit our use of threads to a bounded pool: import threading, time, Queue class MultiThread(object): def _ _init_ _(self, function, argsVector, maxThreads=5, queue_results=False): self._function = function self._lock = threading.Lock( ) self._nextArgs = iter(argsVector).next self._threadPool = [ threading.Thread(target=self._doSome) for i in range(maxThreads) ] if queue_results: self._queue = Queue.Queue( ) else: self._queue = None def _doSome(self): while True: self._lock.acquire( ) try: try: args = self._nextArgs( ) except StopIteration: break finally: self._lock.release( ) result = self._function(args) if self._queue is not None: self._queue.put((args, result)) def get(self, *a, **kw): if self._queue is not None: return self._queue.get(*a, **kw) else: raise ValueError, 'Not queueing results' def start(self): for thread in self._threadPool: time.sleep(0) # necessary to give other threads a chance to run thread.start( ) def join(self, timeout=None): for thread in self._threadPool: thread.join(timeout) if _ _name_ _=="_ _main_ _": import random def recite_n_times_table(n): for i in range(2, 11): print "%d * %d = %d" % (n, i, n * i) time.sleep(0.3 + 0.3*random.random( )) mt = MultiThread(recite_n_times_table, range(2, 11)) mt.start( ) mt.join( ) print "Well done kids!" Discussion This recipe's MultiThread class offers a simple way to execute a function in parallel, on many sets of arguments, using a bounded pool of threads. Optionally, you can ask for results of the calls to the function to be queued, so you can retrieve them, but by default the results are just thrown away. The MultiThread class takes as its arguments a function, a sequence of argument tuples for said function, and optionally a boundary on the number of threads to use in its pool and an indicator that results should be queued. Beyond the constructor, it exposes three methods: start, to start all the threads in the pool and begin the parallel evaluation of the function over all argument tuples; join, to perform a join on all threads in the pool (meaning to wait for all the threads in the pool to have terminated); and get, to get queued results (if it was instantiated with the optional flag queue_results set to TRue, to ask for results to be queued). Internally, class MultiThread uses its private method doSome as the target callable for all threads in the pool. Each thread works on the next available tuple of arguments (supplied by the next method of an iterator on the iterable whose items are such tuples, with the call to next being guarded by the usual locking idiom), until all work has been completed. As is usual in Python, the module can also be run as a free-standing main script, in which case it runs a simple demonstration and self-test. In this case, the demonstration simulates a class of schoolchildren reciting multiplication tables as fast as they can. Real use cases for this recipe mostly involve functions that are I/O bound, meaning functions that spend substantial time performing I/O. If a function is "CPU bound", meaning the function spends its time using the CPU, you get better overall performance by performing the computations one after the other, rather than in parallel. In Python, this observation tends to hold even on machines that dedicate multiple CPUs to your program, because Python uses a GIL (Global Interpreter Lock), so that pure Python code from a single process does not run simultaneously on more than one CPU at a time. Input/output operations release the GIL, and so can (and should) any C-coded Python extension that performs substantial computations without callbacks into Python. So, it is possible that parallel execution may speed up your program, but only if either I/O or a suitable C-coded extension is involved, rather than pure computationally intensive Python code. (Implementations of Python on different virtual machines, such as Jython, which runs on a JVM [Java Virtual Machine], or IronPython, which runs on the Microsoft .NET runtime, are of course not bound by these observations: these observations apply only to the widespread "classical Python", meaning CPython, implementation.) See Also Library Reference and Python in a Nutshell docs on modules threading and Queue. Recipe 9.6. Coordinating Threads by Simple Message Passing Credit: Michael Hobbs Problem You want to write a multithreaded application, using, as the synchronization and communication primitive, a simple yet powerful message-passing paradigm. Solution The candygram module lets you use concurrent programming semantics that are essentially equivalent to those of the Erlang language. To use candygram, you start by defining appropriate classes, such as the following one, to model your threads' functionality: import candygram as cg class ExampleThread(object): """A thread-class with just a single counter value and a stop flag.""" def _ _init_ _(self): """ Initialize the counter to 0, the running-flag to True. """ self.val = 0 self.running = True def increment(self): """ Increment the counter by one. """ self.val += 1 def sendVal(self, msg): """ Send current value of counter to requesting thread. """ req = msg[0] req.send((cg.self( ), self.val)) def setStop(self): """ Set the running-flag to False. """ self.running = False def run(self): """ The entry point of the thread. """ # Register the handler functions for various messages: r = cg.Receiver( ) r.addHandler('increment', self.increment) r.addHandler((cg.Process, 'value'), self.sendVal, cg.Message) r.addHandler('stop', self.setStop) # Keep handling new messages until a stop has been requested while self.running: r.receive( ) To start a thread running this code under candygram, use: counter = cg.spawn(ExampleThread( ).run) To handle the counter tHRead's responses, you need another Receiver object, with the properhandler registered: response = cg.Receiver( ) response.addHandler((counter, int), lambda msg: msg[1], cg.Message) And here is an example of how you might use these counter and response objects: # Tell thread to increment twice counter.send('increment') counter.send('increment') # Request the thread's current value, then print the thread's response counter.send((cg.self( ), 'value')) print response.receive( ) # Tell thread to increment one more time counter.send('increment') # Again, request the thread's current value, then print the thread's response counter.send((cg.self( ), 'value')) print response.receive( ) # Tell the thread to stop running counter.send('stop') Discussion With the candygram module (http://candygram.sourceforge.net), Python developers can send and receive messages between threads using semantics nearly identical to those introduced in the Erlang language (http://www.erlang.org). Erlang is widely respected for its elegant built-in facilities for concurrent programming. Erlang's approach is simple and yet powerful. To communicate with another thread, simply send a message to it. You do not need to worry about locks, semaphores, mutexes, and other such primitives, to share information among concurrent tasks. Developers of multitasking software mostly use message passing only to implement a producer/consumer model. When you combine message passing with the flexibility of a Receiver object, however, it becomes much more powerful. For example, by using timeouts and message patterns, a thread may easily handle its messages as a state machine, or as a priority queue. For those who wish to become more familiar with Erlang, http://www.erlang.org/download/erlang-book-part1.pdf (Concurrent Programming in Erlang) provides a very complete introduction. In particular, the candygram module implements all of the functions described in Chapter 5 and sections 7.2, 7.3, and 7.5 of that book. This recipe offers a very elementary demonstration of how messages are passed between threads using candygram. When you run this recipe as a script, the print statements will output the values 2 and then 3. It's important to understand how the candygram.Receiver class works. The addHandler method requires at least two parameters: the first is a message pattern and the second is a handler function. The Receiver.receive method invokes a registered handler function, and returns that function's result, whenever it finds a message that matches the associated pattern. Any parameters optionally passed to addHandler beyond the first two get passed as parameters to the handler function when the Receiver calls it. If a parameter is the candygram.Message constant, then receive replaces that parameter with the matching message when it calls the handler function. This recipe's code contains four different message patterns: 'increment', (cg.Process, 'value'), 'stop', and (counter, int). The 'increment' and 'stop' patterns are simple patterns that match any message that consists solely of the strings 'increment' and 'stop', respectively. The (cg.Process, 'value') pattern matches any message that is a tuple with two items, where thefirst item isinstance of cg.Process and the second item is the string value. Lastly, the (counter, int) pattern matches any message that is a tuple with two items where the first item is the counter object and the second element is an integer. You can find more information about the Candygram package at http://candygram.sourceforge.net. At that URL, you can find all details on how to specify message patterns, how to set a timeout for the Receiver.receive method, and how to monitor the running status of spawned threads. See Also Concurrent Programming in Erlang at http://www.erlang.org/download/erlang-book-part1.pdf; the candygram home page at http://candygram.sourceforge.net. Recipe 9.7. Storing Per-Thread Information Credit: John E. Barham, Sami Hangaslammi, Anthony Baxter Problem You need to allocate to each thread some storage that only that thread can use. Solution Thread-specific storage is a useful design pattern, and Python 2.3 did not yet support it directly. However, even in 2.3, we could code it up in terms of a dictionary protected by a lock. For once, it's slightly more general, and not significantly harder, to program to the lower-level thread module, rather than to the more commonly useful, higher-level tHReading module that Python offers on top of it: _tss = { } try: import thread except ImportError: # We're running on a single-threaded platform (or, at least, the Python # interpreter has not been compiled to support threads), so we just return # the same dict for every call -- there's only one thread around anyway! def get_thread_storage( ): return _tss else: # We do have threads; so, to work: _tss_lock = thread.allocate_lock( ) def get_thread_storage( ): """ Return a thread-specific storage dictionary. """ thread_id = thread.get_ident( ) _tss_lock.acquire( ) try: return _tss.set_default(thread_id, { }) finally: _tss_lock.release( ) Python 2.4 offers a much simpler and faster implementation, of course, thanks to the new tHReading.local function: try: import threading except ImportError: import dummy_threading as threading _tss = threading.local( ) def get_thread_storage( ): return _tss._ _dict_ _ DiscussionThe main benefit of multithreaded programs is that all of the threads can share global objects when they need to do so. Often, however, each thread also needs some storage of its ownfor example, to store a network or database connection unique to itself. Indeed, each such externally oriented object is generally best kept under the control of a single thread, to avoid multiple possibilities of highly peculiar behavior, race conditions, and so on. The get_thread_storage function in this recipe solves this problem by implementing the "thread- specific storage" design pattern, and specifically by returning a thread-specific storage dictionary. The calling thread can then use the returned dictionary to store any kind of data that is private to the thread. This recipe is, in a sense, a generalization of the get_transaction function from ZODB, the object-oriented database underlying Zope. One possible extension to this recipe is to add a delete_thread_storage function. Such a function would be useful, particularly if a way could be found to automate its being called upon thread termination. Python's threading architecture does not make this task particularly easy. You could spawn a watcher thread to do the deletion after a join with the calling thread, but that's a rather heavyweight approach. The recipe as presented, without deletion, is quite appropriate for the common and recommended architecture in which you have a pool of (typically daemonic) worker threads (perhaps some of them general workers, with others dedicated to interfacing to specific external resources) that are spawned at the start of the program and do not go away until the end of the whole process. When multithreading is involved, implementation must always be particularly careful to detect and prevent race conditions, deadlocks, and other such conflicts. In this recipe, I have decided not to assume that a dictionary's set_default method is atomic (meaning that no thread switch can occur while set_default executes)adding a key can potentially change the dictionary's whole structure, after all. If I was willing to make such an assumption, I could do away with the lock and vastly increase performance, but I suspect that such an assumption might make the code too fragile and dependent on specific versions of Python. (It seems to me that the assumption holds for Python 2.3, but, even if that is the case, I want my applications to survive subtle future changes to Python's internals.) Another risk is that, if a thread terminates and a new one starts, the new thread might end up with the same thread ID as the just-terminated one, and therefore accidentally share the "thread-specific storage" dictionary left behind by the just-terminated thread. This risk might be mitigated (though not eliminated) by providing the delete_thread_storage function mentioned in the previous paragraph. Again, this specific problem does not apply to me, given the kind of multithreading architecture that I use in my applications. If your architecture differs, you may want to modify this recipe's solution accordingly. If the performance of this recipe's version is insufficient for your application's needs, due to excessive overhead in acquiring and releasing the lock, then, rather than just removing the lock at the risk of making your application fragile, you might consider an alternative: _creating_threads = True _tss_lock = thread.allocate_lock( ) _tss = { } class TssSequencingError(RuntimeError): pass def done_creating_threads( ): """ switch from thread-creation to no-more-threads-created state """ global _creating_threads if not _creating_threads: raise TssSequencingError('done_creating_threads called twice') _creating_threads = False def get_thread_storage( ): """ Return a thread-specific storage dictionary. """ thread_id = thread.get_ident( ) # fast approach if thread-creation phase is finished if not _creating_threads: return _tss[thread_id] # careful approach if we're still creating threads try: _tss_lock.acquire( ) return _tss.setdefault(thread_id, { }) finally: _tss_lock.release( ) This variant adds a boolean switch _creating_threads, initially true. As long as the switch is true, the variant uses a careful locking-based approach, quite similar to the one presented in this recipe's Solution. At some point in time, when all threads that will ever exist (or at least all that will ever require access to get_thread_storage) have been started, and each of them has obtained its thread-local storage dictionary, your application calls done_creating_threads. This sets _creating_threads to False, and every future call to get_thread_storage then takes a fast path where it simply indexes into global dictionary _tssno more acquiring and releasing the lock, no more creating a thread's thread-local storage dictionary if it didn't yet exist. As long as your application can determine a moment in which it can truthfully call done_creating_threads, the variant in this subsection should definitely afford a substantial increase in speed compared to this recipe's Solution. Note that it is particularly likely that you can use this variant if your application follows the popular and recommended architecture mentioned previously: a bounded set of daemonic, long-lived worker threads, all created early in your program. This is fortunate, because, if your application is performance-sensitive enough to worry about the locking overhead of this recipe's solution, then no doubt you will want to structure your application that way. The alternative approach of having many short-lived threads is generally quite damaging to performance. If your application needs to run only under Python 2.4, you can get a much simpler, faster, and solid implementation by relying on the new threading.local function. threading.local returns a new object on which any thread can get and set arbitrary attributes, independently from whatever getting and setting other threads may be doing on the same object. This recipe, in the 2.4 variant, returns the per-thread _ _dict_ _ of such an object, for uniformity with the 2.3 variant. This way, your applications can be made to run on both Python 2.3 and 2.4, using the best version in each case: import sys if sys.version >= '2.4': # insert 2.4 definition of get_local_storage here else: # insert 2.3 definition of get_local_storage here The 2.4 variant of this recipe also shows off the intended use of module dummy_threading, which, like its sibling dummy_thread, is also available in Python 2.3. By conditionally using these dummy modules, which are available on all platforms, whether or not Python was compiled with thread support, you may sometimes, with due care, be able to write applications that can run on any platform, taking advantage of threading where it's available but running anyway even where threading is not available. In the 2.3 variant, we did not use the similar approach based on dummy_thread, because the overhead would be too high to pay on nonthreaded platforms; in the 2.4 variant, overhead is pretty low anyway, so we went for the simplicity that dummy_threading affords. See Also For an exhaustive treatment of the design pattern that describes thread-specific storage (albeit aimed at C++ programmers), see Douglas Schmidt, Timothy Harrisson, Nat Pryce, Thread- Specific Storage: An Object Behavioral Pattern for Efficiently Accessing per-Thread State (http://www.cs.wustl.edu/~schmidt/PDF/TSS-pattern.pdf); the Library Reference documentation dummy_thread, dummy_threading, and Python 2.4's tHReading.local; ZODB at http://zope.org/Wikis/ZODB/FrontPage. Recipe 9.8. Multitasking Cooperatively Without Threads Credit: Brian Bush, Troy Melhase, David Beach, Martin Miller Problem You have a task that seems suited to multithreading, but you don't want to incur the overhead that real thread-switching would impose. Solution Generators were designed to simplify iteration, but they're also quite suitable as a basis for cooperative multitasking, also known as microthreading: import signal # credit: original idea was based on an article by David Mertz # http://gnosis.cx/publish/programming/charming_python_b7.txt # some example 'microthread' generators def empty(name): """ This is an empty task for demonstration purposes. """ while True: print "", name yield None def terminating(name, maxn): """ This is a counting task for demonstration purposes. """ for i in xrange(maxn): print "Here %s, %s out of %s" % (name, i, maxn) yield None print "Done with %s, bailing out after %s times" % (name, maxn) def delay(duration=0.8): """ Do nothing at all for 'duration' seconds. """ import time while True: print "" % duration time.sleep(duration) yield None class GenericScheduler(object): def _ _init_ _(self, threads, stop_asap=False): signal.signal(signal.SIGINT, self.shutdownHandler) self.shutdownRequest = False self.threads = threads self.stop_asap = stop_asap def shutdownHandler(self, n, frame): """ Initiate a request to shutdown cleanly on SIGINT.""" print "Request to shut down." self.shutdownRequest = True def schedule(self): def noop( ): while True: yield None n = len(self.threads) while True: for i, thread in enumerate(self.threads): try: thread.next( ) except StopIteration: if self.stop_asap: return n -= 1 if n==0: return self.threads[i] = noop( ) if self.shutdownRequest: return if _ _name_ _== "_ _main_ _": s = GenericScheduler([ empty('boo'), delay( ), empty('foo'), terminating('fie', 5), delay(0.5), ], stop_asap=True) s.schedule( ) s = GenericScheduler([ empty('boo'), delay( ), empty('foo'), terminating('fie', 5), delay(0.5), ], stop_asap=False) s.schedule( ) Discussion Microthreading (or cooperative multitasking) is an important technique. If you want to pursue it in earnest for complex uses, you should definitely look up the possibilities of Christian Tismer's Stackless, a Python version specialized for microthreading, at http://www.stackless.com/. However, you can get a taste of cooperative multitasking without straying from Python's core, by making creative use of generators, as shown in this recipe. A simple approach to cooperative multitasking, such as the one presented in this recipe, is not suitable when your tasks must perform long-running work, particularly I/O tasks that may involve blocking system calls. For such applications, look into real threading, or, as a strong alternative, look into the event-driven approach offered by module asyncore in the Python Standard Library (on a simple scale) and by package Twisted at http://twistedmatrix.com/products/twisted (on a grandiose scale). But if your application has modest I/O needs, and you can slice up any computation your tasks perform into short chunks, each of which you can end with a yield, this recipe may be just what you're looking for. See Also David Mertz's site, chock-full of idiosyncratic, fascinating ideas, is at http://gnosis.cx/; Christian Tismer's Stackless Python, the best way to do cooperative multitasking in Python (and much else besides), is at http://www.stackless.com/; Twisted Matrix, the best way to do event-driven (asynchronous) programming, is at http://twistedmatrix.com/. Recipe 9.9. Determining Whether Another Instanceof a Script Is Already Running in Windows Credit: Bill Bell Problem In a Windows environment, you want to ensure that only one instance of a script is running at any given time. Solution Many tricks can be used to avoid starting multiple copies of an application, but they're all quite fragileexcept those based on a mutual-exclusion (mutex) kernel object, such as this one. Mark Hammond's precious PyWin32 package supplies all the needed hooks into the Windows APIs to let us exploit a mutex for this purpose: from win32event import CreateMutex from win32api import GetLastError from winerror import ERROR_ALREADY_EXISTS from sys import exit handle = CreateMutex(None, 1, 'A unique mutex name') if GetLastError( ) == ERROR_ALREADY_EXISTS: # Take appropriate action, as this is the second # instance of this script; for example: print 'Oh! dear, I exist already.' exit(1) else: # This is the only instance of the script; let # it do its normal work. For example: from time import sleep for i in range(10): print "I'm running",i sleep(1) Discussion The string 'A unique mutex name' must be chosen to be unique to this script, and it must not be dynamically generated, because the string must have the same value for all potential simultaneous instances of the same script. A fresh, globally unique ID that you manually generate and insert at script-authoring time would be a good choice. According to the Windows documentation, the string can contain any characters except backslashes (\). On Windows platforms that implement Terminal Services, you can optionally prefix the string with Global\ or Local\, but such prefixes would make the string invalid for most versions of Windows, including NT, 95, 98, and ME. The Win32 API call CreateMutex creates a Windows kernel object of the mutual-exclusion (mutex) kind and returns a handle to it. Note that we do not close this handle, because it needs to exist throughout the time this process is running. It's important to let the Windows kernel take care of removing the handle (and the object it indicates, if the handle being removed is the onlyhandle to that kernel object) when our process terminates. The only thing we really care about is the return code from the API call, which we obtain by calling the GetLastError API right after it. That code is ERROR_ALREADY_EXISTS if and only if the mutual-exclusion object we tried to create already exists (i.e., if another instance of this script is already running). This approach is perfectly safe and not subject to race conditions and similar anomalies, even if two instances of the script are trying to start at the same time (a reasonably frequent occurrence, e.g., if the user erroneously double-clicks in an Active Desktop setting where a single click already starts the application). The Windows specifications guarantee that only one of the instances will create the mutex, while the other will be informed that the mutex already exists. Mutual exclusion is therefore guaranteed by the Windows kernel itself, and the recipe is as solid as the operating system. See Also Documentation for the Win32 API in PyWin32 (http://starship.python.net/crew/mhammond/win32/Downloads.html) or ActivePython (http://www.activestate.com/ActivePython/); Windows API documentation available from Microsoft (http://msdn.microsoft.com); Python Programming on Win32, by Mark Hammond and Andy Robinson (O'Reilly). Recipe 9.10. Processing Windows Messages Using MsgWaitForMultipleObjects Credit: Michael Robin Problem In a Win32 application, you need to process messages, but you also want to wait for kernel-level waitable objects, and coordinate several activities. Solution A Windows application's message loop, also known as its message pump, is at the heart of Windows. It's worth some effort to ensure that the heart beats properly and regularly: import win32event import pythoncom TIMEOUT = 200 # ms StopEvent = win32event.CreateEvent(None, 0, 0, None) OtherEvent = win32event.CreateEvent(None, 0, 0, None) class myCoolApp(object): def OnQuit(self): # assume 'areYouSure' is a global function that makes a final # check via a message box, a fancy dialog, or whatever else! if areYouSure( ): win32event.SetEvent(StopEvent) # Exit msg pump def _MessagePump( ): waitables = StopEvent, OtherEvent while True: rc = win32event.MsgWaitForMultipleObjects( waitables, , # Wait for all = false, so it waits for any one TIMEOUT, # (or win32event.INFINITE) win32event.QS_ALLEVENTS) # Accept all kinds of events # You can call a function here, if it doesn't take too long. It will # be executed at least every TIMEOUT ms -- possibly a lot more often, # depending on the number of Windows messages received. if rc == win32event.WAIT_OBJECT_0: # Our first event listed, the StopEvent, was triggered, so # we must exit, terminating the message pump break elif rc == win32event.WAIT_OBJECT_0+1: # Our second event listed, "OtherEvent", was set. Do # whatever needs to be done -- you can wait on as many # kernel-waitable objects as needed (events, locks, # processes, threads, notifications, and so on). pass elif rc == win32event.WAIT_OBJECT_0+len(waitables): # A windows message is waiting - take care of it. (Don't # ask me why a WAIT_OBJECT_MSG isn't defined < # WAIT_OBJECT_0...!). # This message-serving MUST be done for COM, DDE, and other # Windows-y things to work properly! if pythoncom.PumpWaitingMessages( ): break # we received a wm_quit message elif rc == win32event.WAIT_TIMEOUT: # Our timeout has elapsed. # Do some work here (e.g, poll something you can't thread) # or just feel good to be alive. pass else: raise RuntimeError("unexpected win32wait return value") Discussion Most Win32 applications must process messages, but you often want to wait on kernel waitables and coordinate a lot of things going on at the same time. A good message pump structure is the key to this, and this recipe exemplifies a reasonably simple but pretty effective one. With the message pump shown in this recipe, messages and other events get dispatched as soon as they are posted, and a timeout allows you to poll other components. You may need to poll if the proper calls or event objects are not exposed in your Win32 event loop, as many components insist on running only on the application's main thread and cannot run on spawned (secondary) threads. You can add many other refinements, just as you can to any other Win32 message-pump approach. Python lets you do this with as much precision as C does, thanks to Mark Hammond's PyWin32 package (which used to be known as win32all). However, the relatively simple message pump presented in this recipe is already a big step up from the typical naive application that can either serve its message loop or wait on kernel waitables, but not both. The key to this recipe is the Windows API call MsgWaitForMultipleObjects, which takes several parameters. The first is a tuple of kernel objects you want to wait for. The second parameter is a flag that is normally 0. The value 1 indicates that you should wait until all the kernel objects in the first parameter are signaled, but my experience suggests that you almost invariably want to stop waiting when any one of these objects is signaled, so this parameter will almost always be 0. The third is a flag that specifies which Windows messages you want to interrupt the wait; always pass win32event.QS_ALLEVENTS here, to make sure any Windows message interrupts the wait. The fourth parameter is a timeout period (in milliseconds), or win32event.INFINITE if you are sure you do not need to do any periodic polling. This function is a polling loop and, sure enough, it loops (with a while True, which is terminated only by a break within it). At each leg of the loop, it calls the API that waits for multiple objects. When that API stops waiting, it returns a code that explains why it stopped waiting. A value between win32event.WAIT_OBJECT_0 and win32event.WAIT_OBJECT_0+N-1 (where N is the number of waitable kernel objects in the tuple you passed as the first parameter), inclusive, means that the wait finished because an object was signaled (being signaled means different things for each kind of waitable kernel object). The return code's difference from win32event.WAIT_OBJECT_0 is the index of the relevant object in the tuple. A return value of win32event.WAIT_OBJECT_0+N means that the wait finished because a message was pending, and in this case, our recipe processes all pending Windows messages via a call to pythoncom.PumpWaitingMessages. (That function, in turn, returns a true result if a WM_QUIT message was received, so in this case, we break out of the whole while loop.) A code of win32event.WAIT_TIMEOUT means the wait finished because of a timeout, so we can do our polling there. In this case, no message is waiting, and none of our kernel objects of interest were signaled. Basically, the way to tune this recipe for yourself is by using the right kernel objects as waitables (with an appropriate response to each) and by doing whatever you need to do periodically in thepolling case. While this means you must have some detailed understanding of Win32, of course, it's still quite a bit easier than designing your own special-purpose, message-loop function from scratch. I suspect that a purist would find some way or other to wrap all of this message pumping into a neat module, letting each application customize its use of the module by passing in a list of waitables, some dictionary to map different waitables to chunks of code to execute, and a partridge in a pear tree. Go ahead, turn it all into a custom metaclass if you wish, see if I care. For once, though, I think the right approach to reusing this code is to copy it into your application's source directories, and use your trusty text editor (gasp!) to tailor the message pump to your application's exact needs. See Also Documentation for the Win32 API in PyWin32 (http://starship.python.net/crew/mhammond/win32/Downloads.html) or ActivePython (http://www.activestate.com/ActivePython/); Windows API documentation available from Microsoft (http://msdn.microsoft.com); Mark Hammond and Andy Robinson, Python Programming on Win32 (O'Reilly). Recipe 9.11. Driving an External Process with popen Credit: Sébastien Keim, Tino Lange, Noah Spurrier Problem You want to drive an external process that accepts commands from its standard input, and you don't care about the responses (if any) that the external process may emit on its standard output. Solution If you need to drive only the other process' input and don't care about its output, the simple os.popen function is enough. For example, here is a way to do animated graphics by driving the free program gnuplot via os.popen: import os f = os.popen('gnuplot', 'w') print >>f, "set yrange[-300:+300]" for n in range(300): print >>f, "plot %i*cos(x)+%i*log(x+10)" % (n, 150-n) f.flush( ) f.close( ) Discussion When you want to use Python as a glue language, sometimes (in particularly easy cases) the simple function popen (from the standard library module os) may be all you need. Specifically, os.popen may suffice when you need to drive an external program that accepts commands on its standard input, as long as you can ignore any response that the program might be making on its standard output (and also error messages that the program might be sending to its standard error). A good example is given by the free plotting program gnuplot. (os.popen may also suffice when you need to obtain the output from a program that does not need to read its standard input.) The statement f = os.popen('gnuplot', 'w') creates a file-like object connected to the standard input of the program it launches, namely 'gnuplot'. (To try this recipe, you have to have gnuplot installed on your PATH, but since gnuplot is freely available and widely ported software, that should not be a problem!) Whatever we write to f, the external process receives on its standard input, just as would happen if we used that same program interactively. For more of the same, check out http://sourceforge.net/projects/gnuplot-py/: it's a rich and interesting Python interface to gnuplot implemented entirely on the basis of the simple idea shown in this recipe! When your needs are more sophisticated than os.popen can accommodate, you may want to look at os.popen2 and other such higher-numbered functions in module os, or, in Python 2.4, the new standard library module subprocess. However, in many cases, you're likely to be disappointed: as soon as you get beyond the basics, driving (from your own programs) other external programs that were designed to be used interactively can become more than a little frustrating. Fortunately, a solution is at hand: it's pexpect, a third-party Python module that you can find at http://pexpect.sourceforge.net/. pexpect is designed specifically for the task ofdriving other programs, and it lets you check on the other program's responses as well as sending commands to the other program's standard input. Still, while pexpect will most definitely offer you all the power you need, os.popen will probably suffice when you don't need anything fancy! See Also Module os (specifically os.popen) in the Library Reference and Python in a Nutshell; gnuplot is at http://www.gnuplot.info/; gnuplot.py is at http://sourceforge.net/projects/gnuplot-py/; pexpect is at http://pexpect.sourceforge.net/. Recipe 9.12. Capturing the Output and Error Streams from a Unix Shell Command Credit: Brent Burley, Bradey Honsinger, Tobias Polzin, Jonathan Cano, Padraig Brady Problem You need to run an external process in a Unix-like environment and capture both the output and error streams from that external process. Solution The popen2 module lets you capture both streams, but you also need help from module fcntl, to make the streams nonblocking and thus avoid deadlocks, and from module select, to orchestrate the action: import os, popen2, fcntl, select def makeNonBlocking(fd): fl = fcntl.fcntl(fd, os.F_GETFL) try: fcntl.fcntl(fd, os.F_SETFL, fl | os.O_NDELAY) except AttributeError: fcntl.fcntl(fd, os.F_SETFL, fl | os.FNDELAY) def getCommandOutput(command): child = popen2.Popen3(command, 1) # Capture stdout and stderr from command child.tochild.close( ) # don't need to write to child's stdin outfile = child.fromchild outfd = outfile.fileno( ) errfile = child.childerr errfd = errfile.fileno( ) makeNonBlocking(outfd) # Don't deadlock! Make fd's nonblocking. makeNonBlocking(errfd) outdata, errdata = [ ], [ ] outeof = erreof = False while True: to_check = [outfd]*(not outeof) + [errfd]*(not erreof) ready = select.select(to_check, [ ], [ ]) # Wait for input if outfd in ready[0]: outchunk = outfile.read( ) if outchunk == '': outeof = True else: outdata.append(outchunk) if errfd in ready[0]: errchunk = errfile.read( ) if errchunk == '': erreof = True else: errdata.append(errchunk) if outeof and erreof: break select.select([ ],[ ],[ ],.1) # Allow a little time for buffers to fill err = child.wait( ) if err != 0: raise RuntimeError, '%r failed with exit code %d\n%s' % ( command, err, ''.join(errdata)) return ''.join(outdata) def getCommandOutput2(command): child = os.popen(command) data = child.read( ) err = child.close( ) if err: raise RuntimeError, '%r failed with exit code %d' % (command, err) Discussion This recipe shows how to execute a Unix shell command and capture the output and error streams in Python. By contrast, os.system sends both streams directly to the terminal. The function getCommandOutput presented in this recipe executes a command and returns the command's output. If the command fails, getCommandOutput raises an exception, using the text captured from the command's stderr as part of the exception's arguments. Most of the complexity of this code is due to the difficulty of capturing both the output and error streams of the child process independently and at the same time. Normal (blocking) read calls may deadlock if the child is trying to write to one stream, and the parent is waiting for data on the other stream; so, the streams must be set to nonblocking, and select must be used to wait for data on either of the streams. Note that the second select call is included just to add a 0.1-second sleep after each read. Counter intuitively, this allows the code to run much faster, since it gives the child time to put more data in the buffer. Without it, the parent may try to read only a few bytes at a time, which can be very expensive. Calling time.sleep(0.1) should be exactly equivalent, but since I was already, necessarily, calling select.select elsewhere in the recipe's code, I decided not to also import module time needlessly. If you want to capture only the output and don't mind the error stream going to the terminal, you can use the much simpler code presented in getCommandOutput2. If you want to suppress the error stream altogether, that's easy, toojust append 2>/dev/null to the command. For example: listing = getCommandOutput2('ls -1 2>/dev/null') Another possibility is given by the os.popen4 function, which combines the output and error streams of the child process. However, in that case the streams are combined in a potentially messy way, depending on how they are buffered in the child process, so this recipe can help. In Python 2.4, you can use class Popen, instead of popen2.Popen3, from the new standard library module subprocess. However, the issues highlighted in this recipe (namely, the need to use modules fcntl and select to make files nonblocking and coordinate the loop that interacts with the child process) aren't really affected by whether you use popen2 or subprocess. This recipe does, as advertised, require a rather Unix-like underlying platform. Cygwin, which does a generally great job of emulating Unix on top of Windows, is not sufficient; for example, it offers no way to set files to nonblocking mode, nor to select on general files. (Under Windows, you are allowed to select only on sockets, not on other files.) If you must run on such problematic, non-Unix platforms, you may prefer a very different approach, based on using temporary files: import os, tempfiledef getCommandOutput(command): outfile = tempfile.mktemp( ) errfile = tempfile.mktemp( ) cmd = "( %s ) > %s 2> %s" % (command, outfile, errfile) err = os.system(cmd) >> 8 try: if err != 0: raise RuntimeError, '%r failed with exit code %d\n%s' % ( command, err, file(errfile).read( )) return file(outfile).read( ) finally: os.remove(outfile) os.remove(errfile) See Also Documentation of the standard library modules os, popen2, fcntl, select, and tempfile in the Library Reference and Python in a Nutshell; (Python 2.4 only) module subprocess in the Library Reference. Recipe 9.13. Forking a Daemon Process on Unix Credit: Jürgen Hermann, Andy Gimblett, Josh Hoyt, Noah Spurrier, Jonathan Bartlett, Greg Stein Problem You need to fork a daemon process on a Unix or Unix-like system, which, in turn, requires a certain precise sequence of system calls. Solution Unix daemon processes must detach from their controlling terminal and process group. Doing so is not hard, but it does require some care, so it's worth writing a daemonize.py module once and for all: import sys, os ''' Module to fork the current process as a daemon. NOTE: don't do any of this if your daemon gets started by inetd! inetd does all you need, including redirecting standard file descriptors; the chdir( ) and umask( ) steps are the only ones you may still want. ''' def daemonize (stdin='/dev/null', stdout='/dev/null', stderr='/dev/null'): ''' Fork the current process as a daemon, redirecting standard file descriptors (by default, redirects them to /dev/null). ''' # Perform first fork. try: pid = os.fork( ) if pid > 0: sys.exit(0) # Exit first parent. except OSError, e: sys.stderr.write("fork #1 failed: (%d) %s\n" % (e.errno, e.strerror)) sys.exit(1) # Decouple from parent environment. os.chdir("/") os.umask(0) os.setsid( ) # Perform second fork. try: pid = os.fork( ) if pid > 0: sys.exit(0) # Exit second parent. except OSError, e: sys.stderr.write("fork #2 failed: (%d) %s\n" % (e.errno, e.strerror)) sys.exit(1) # The process is now daemonized, redirect standard file descriptors. for f in sys.stdout, sys.stderr: f.flush( ) si = file(stdin, 'r') so = file(stdout, 'a+') se = file(stderr, 'a+', 0) os.dup2(si.fileno( ), sys.stdin.fileno( )) os.dup2(so.fileno( ), sys.stdout.fileno( )) os.dup2(se.fileno( ), sys.stderr.fileno( )) def _example_main ( ): ''' Example main function: print a count & timestamp each second ''' import time sys.stdout.write('Daemon started with pid %d\n' % os.getpid( ) ) sys.stdout.write('Daemon stdout output\n') sys.stderr.write('Daemon stderr output\n') c = 0 while True: sys.stdout.write('%d: %s\n' % (c, time.ctime( ))) sys.stdout.flush( ) c = c + 1 time.sleep(1) if _ _name_ _ == "_ _main_ _": daemonize('/dev/null','/tmp/daemon.log','/tmp/daemon.log') _example_main( ) Discussion Forking a daemon on Unix requires a certain specific sequence of system calls, which is explained in W. Richard Stevens' seminal book, Advanced Programming in the Unix Environment (Addison- Wesley). We need to fork twice, terminating each parent process and letting only the grandchild of the original process run the daemon's code. This allows us to decouple the daemon process from the calling terminal, so that the daemon process can keep running (typically as a server process without further user interaction, like a web server) even after the calling terminal is closed. The only visible effect of doing so is that when your script runs this module's daemonize function, you get your shell prompt back immediately. For all of the details about how and why this works in Unix and Unix-like systems, see Stevens' wonderful book. Another important source of information on both practical and theoretical issues about "daemon forking" can be found as part of the Unix Programming FAQ, at http://www.erlenstar.demon.co.uk/unix/faq_2.html#SEC16. To summarize: the first fork lets the shell return, and also lets you do a setsid (to remove you from your controlling terminal, so you can't accidentally be sent a signal). However, setsid makes this process a "session leader", which means that if the process ever opens any terminal, it will become the process' controlling terminal. We do not want a daemon to have any controlling terminal, which is why we fork again. After the second fork, the process is no longer a "session leader", so it can open any file (including a terminal) without thereby accidentally reacquiring a controlling terminal. Both Stevens and the Unix Programming FAQ provide examples in the C programming language, but since the Python Standard Library exposes a full POSIX interface, you can also do it all in Python. Typical C code for a daemon fork translates almost literally to Python; the only difference you have to care abouta minor detailis that Python's os.fork does not return -1 on errors, but rather throws an OSError exception. Therefore, rather than testing for a less-than-zero return code from fork, as we would in C, we run the fork in the try clause of a TRy/except statement, so that we can catch the exception, should it happen, and print appropriate diagnostics to standard error. See Also Documentation of the standard library module os in the Library Reference and Python in a Nutshell; Unix manpages for the fork, umask, and setsid system calls; W.Richard Stevens, Advanced Programming in the Unix Environment (Addison-Wesley); also, the Unix Programming FAQ on daemon forking, at http://www.erlenstar.demon.co.uk/unix/faq_2.html#SEC16. Chapter 10. System Administration Introduction Recipe 10.1. Generating Random Passwords Recipe 10.2. Generating Easily Remembered Somewhat-Random Passwords Recipe 10.3. Authenticating Users by Means of a POP Server Recipe 10.4. Calculating Apache Hits per IP Address Recipe 10.5. Calculating the Rate of Client Cache Hits on Apache Recipe 10.6. Spawning an Editor from a Script Recipe 10.7. Backing Up Files Recipe 10.8. Selectively Copying a Mailbox File Recipe 10.9. Building a Whitelist of Email Addresses From a Mailbox Recipe 10.10. Blocking Duplicate Mails Recipe 10.11. Checking Your Windows Sound System Recipe 10.12. Registering or Unregistering a DLL on Windows Recipe 10.13. Checking and Modifying the Set of Tasks Windows Automatically Runs at Login Recipe 10.14. Creating a Share on Windows Recipe 10.15. Connecting to an Already Running Instance of Internet Explorer Recipe 10.16. Reading Microsoft Outlook Contacts Recipe 10.17. Gathering Detailed System Informationon Mac OS X Introduction Credit: Donn Cave, University of Washington In this chapter, we consider a class of programmerthe humble system administratorin contrast to other chapters' focus on functional domains. As a programmer, the system administrator faces most of the same problems that other programmers face and should find the rest of this book of at least equal interest. Python's advantages in the system administration domain are also quite familiar to other Python programmers, but Python's competition is different. On Unix platforms, at any rate, the landscape is dominated by a handful of lightweight languages such as the Bourne shell and awk that aren't exactly made obsolete by Python. These little languages can often support a simpler, clearer, and more concise solution than Python, particularly for commands that you're typing interactively at the shell command prompt. But Python can do things these languages can't, and it's often more robust when dealing with issues such as unusually large data inputs. Another notable competitor, especially on Unix systems, is Perl (which isn't really a little language at all), with just about the same overall power as Python, and usable for typing a few commands interactively at the shell's command prompt. Python's strength here is readability and maintainability: when you dust off a script you wrote in a hurry eight months ago, because you need to make some changes to it, you don't spend an hour to figure out whatever exactly you had in mind when you wrote this or that subtle trick. You just don't use any tricks at all, subtle or gross, so that your Python scrips work just fine and you don't burn your time, months later, striving to reverse-engineer them for understanding. One item that stands out in this chapter's solutions is the wrapper: the alternative, programmed interface to a software system. On Unix (including, these days, Mac OS X), this is usually a fairly prosaic matter of diversion and analysis of text I/O. Life is easy when the programs you're dealing with are able to just give clean textual output, without requiring complex interaction (see Eric Raymond, The Art of Unix Programming, http://www.faqs.org/docs/artu/, for an informative overview of how programs should be architected to make your life easy). However, even when you have to wrap a program that's necessarily interactive, all is far from lost. Python has very good support in this area, thanks, first of all, to the fact that it places C-level pseudo-TTY functions at your disposal (see the pty module of the Python Standard Library). The pseudo-TTY device is like a bidirectional pipe with TTY driver support, so it's essential for things such as password prompts that insist on a TTY. Because it appears to be a TTY, applications writing to a pseudo-TTY normally use line buffering, instead of the block buffering that gives problems with pipes. Pipes are more portable and less trouble to work with, but they don't work for interfacing to every application. Excellent third-party extensions exist that wrap pty into higher-level layers for ease of use, most notably Pexpect, http://pexpect.sourceforge.net/. On Windows, the situation often is not as prosaic as on Unix-like platforms, since the information you need to do your system administration job may be somewhere in the registry, may be available via some Windows APIs, and/or may be available via COM. The standard Python library _winreg module, Mark Hammond's PyWin32 package, and Thomas Heller's ctypes, taken together, give the Windows administrator reasonably easy access to all of these sources, and you'll see more Windows administration recipes here than you will ones for Unix. The competition for Python as a system administration language on Windows is feeble compared to that on Unix, which is yet another reason for the platform's prominence here. The PyWin32 extensions are available for download at http://sourceforge.net/projects/pywin32/. PyWin32 also comes with ActiveState's ActivePython distribution of Python (http://www.activestate.com/ActivePython/). To use this rich and extremely useful package most effectively, you also need Mark Hammond and Andy Robinson, Python Programming on Win32 (O'Reilly). ctypes is available for download at http://sourceforge.net/projects/ctypes. While it may sometimes be difficult to see what brought all the recipes together in this chapter, itisn't difficult to see why system administrators deserve their own chapter: Python would be nowhere without them! Who else, back when Python was still an obscure, fledgling language, could bring it into an organization and almost covertly infiltrate it into the working environment? If it weren't for the offices of these benevolent and pragmatic anarchists, Python might well have languished in obscurity despite its merits. Recipe 10.1. Generating Random Passwords Credit: Devin Leung Problem You need to create new passwords randomlyfor example, to assign them automatically to new user accounts. Solution One of the chores of system administration is installing new user accounts. Assigning a different, totally random password to each new user is a good idea. Save the following code as makepass.py: from random import choice import string def GenPasswd(length=8, chars=string.letters+string.digits): return ''.join([ choice(chars) for i in range(length) ]) Discussion This recipe is useful when you are creating new user accounts and assigning each of them a different, totally random password. For example, you can print six passwords of length 12: >>> import makepass >>> for i in range(6): ... print makepass.GenPasswd(12) ... uiZWGSJLWjOI FVrychdGsAaT CGCXZAFGjsYI TPpQwpWjQEIi HMBwIvRMoIvh Of course, such totally random passwords, while providing an excellent theoretical basis for security, are impossibly hard to remember for most users. If you require users to stick with their assigned passwords, many users will probably write them down. The best you can hope for is that new users will set their own passwords at their first login, assuming, of course, that the system you're administering lets each user change his own password. (Most operating systems do, but you might be assigning passwords for other kinds of services that unfortunately often lack such facilities.) A password that is written down anywhere is a serious security risk: pieces of paper get lost, misplaced, and peeked at. From a pragmatic point of view, you might be better off assigning passwords that are not totally random; users are more likely to remember them and less likely to write them down (see Recipe 10.2). This practice may violate the theory of password security, but, as all practicing system administrators know, pragmatism trumps theory.See Also Recipe 10.2; documentation of the standard library module random in the Library Reference and Python in a Nutshell. Recipe 10.2. Generating Easily Remembered Somewhat- Random Passwords Credit: Luther Blissett Problem You need to create new passwords randomlyfor example, to assign them automatically to new user accounts. You want the passwords to be somewhat feasible to remember for typical users, so they won't be written down. Solution We can use a pastiche approach for this, mimicking letter n-grams in actual English words. A grander way to look at the same approach is to call it a Markov Chain Simulation of English: import random, string class password(object): # Any substantial file of English words will do just as well: we # just need self.data to be a big string, the text we'll pastiche data = open("/usr/share/dict/words").read( ).lower( ) def renew(self, n, maxmem=3): ''' accumulate into self.chars `n' random characters, with a maximum-memory "history" of `maxmem` characters back. ''' self.chars = [ ] for i in range(n): # Randomly "rotate" self.data randspot = random.randrange(len(self.data)) self.data = self.data[randspot:] + self.data[:randspot] # Get the n-gram where = -1 # start by trying to locate the last maxmem characters in # self.chars. If i1: dopass = int(sys.argv[1]) else: dopass = 8 if len(sys.argv)>2: length = int(sys.argv[2]) else: length = 10 if len(sys.argv)>3: memory = int(sys.argv[3]) else: memory = 3 onepass = password( ) for i in range(dopass): onepass.renew(length, memory) print onepass Discussion This recipe is useful when creating new user accounts and assigning each user a different, random password: it uses passwords that a typical user will find it feasible to remember, hopefully so they won't get written down. See Recipe 10.1 if you prefer totally random passwords. The recipe's idea is based on the good old pastiche concept. Each letter (always lowercase) in the password is chosen pseudo-randomly from data that is a collection of words in a natural language familiar to the users. This recipe uses the file that is /usr/share/dict/words supplied with Linux systems (on my machine, a file of over 45,000 words), but any large document in plain text will do just as well. The trick that makes the passwords sort of memorable, and not fully random, is that each letter is chosen based on the last few letters already picked for the password as it stands so far. Thus, letter transitions will tend to be "repetitive" according to patterns that are familiar to the user. The code in the recipe takes some care to locate each time a random occurrence, in the text being pastiched, of the last maxmem characters picked so far. Since it's easy to find the first occurrence of a substring, the code "rotates" the text string randomly, to ensure that the first occurrence is a random one from the point of view of the original text. If the substring made up with the last maxmem characters picked is not found in the text, the code "backs down" to search for just the last maxmem-1, and so on, backing down until, worst case, it just picks the first character in the rotated text (which is a random character from the point of view of the original text). A break in this Markov Chain process occurs when this picking procedure chooses a character that is not a lowercase letter, in which case, a random lowercase letter is chosen instead (any lowercase letter is picked with equal probability). Here are a couple of typical sample runs of this pastiche.py password-generation script: [situ@tioni cooker]$ python pastiche.py yjackjaceh ackjavagef aldsstordb dingtonous stictlyoke cvaiwandga lidmanneck olexnarinl [situ@tioni cooker]$ python pastiche.py ptiontingt punchankin cypresneyf sennemedwainingrated fancejacev sroofcased nryjackman As you can see, some of these are definitely word-like, others less so, but for a typical human being, none are more problematic to remember than a sequence of even fewer totally random, uncorrelated letters. No doubt some theoretician will complain (justifiably, in a way) that they aren't as random as all that. Well, tough. My point is that they had better not be, if some poor fellow is going to have to remember them! You can compensate for this limitation by making them a bit longer. If said theoretician demonstrates how to compute the entropy per character of this method of password generation (versus the obvious 4.7 bits/character, the base-2 logarithm of 26, for passwords made up of totally random lowercase letters), now that would be a useful contribution indeed. Meanwhile, I'll keep generating passwords this way, rather than in a totally random way. If nothing else, it's the closest thing I've found to a useful application for the lovely pastiche concept. The concept of passwords that are not totally random, but rather a bit more memorable, goes back a long wayat least to the 1960s and to works by Morrie Gasser and Daniel Edwards. A Federal Information Processing Standard (FIPS), FIPS 181, specifies in detail how "pronounceable" passwords are to be generated; see http://www.itl.nist.gov/fipspubs/fip181.htm. See Also Recipe 10.1; documentation of the standard library module random in the Library Reference and Python in a Nutshell. Recipe 10.3. Authenticating Users by Means of a POP Server Credit: Magnus Lyckå Problem You are writing a Python application that must authenticate users. All of the users have accounts on some POP servers, and you'd like to reuse, for your own authentication, the user IDs and passwords that your users have on those servers. Solution To log into the application, a user must provide the server, user ID and password for his mail account. We try logging into that POP server with these credentialsif that attempt succeeds, then the user has authenticated successfully. (Of course, we don't peek into the user's mailbox!) def popauth(popHost, user, passwd): """ Log in and log out, only to verify user identity. Raise exception in case of failure. """ import poplib try: pop = poplib.POP3(popHost) except: raise RuntimeError("Could not establish connection " "to %r for password check" % popHost) try: # Log in and perform a small sanity check pop.user(user) pop.pass_(passwd) length, size = pop.stat( ) assert type(length) == type(size) == int pop.quit( ) except: raise RuntimeError("Could not verify identity. \n" "User name %r or password incorrect." % user) pop.quit( ) Discussion To use this recipe, the application must store somewhere the list of known users and either the single POP server they all share, or the specific POP server on which each user authenticatesit need not be the same POP server for all users. Either a text file, or a simple table in any kind of database, will do just fine for this purpose. This solution is neat, but it does have some weaknesses: Users must trust that any application implementing this authentication system won't abusetheir email accounts. POP passwords are, alas!, sent in plain text over the Internet. We have to trust that the POP server security isn't compromised. Logging in might take a few seconds if the POP server is slow. Logging in won't work if the POP server is down. However, to offset all of these potential drawbacks is the convenience of applications not having to store any passwords, nor forcing a poor overworked system administrator to administer password changes. It's also quite simple! In short, I wouldn't use this approach for a bank system, but I would have no qualms using it, for example, to give users rights to edit web pages at a somewhat restricted WikiWiki, or similarly low-risk applications. See Also Documentation of the standard library module poplib in the Library Reference and Python in a Nutshell. Recipe 10.4. Calculating Apache Hits per IP Address Credit: Mark Nenadov, Ivo Woltring Problem You need to examine a log file from Apache to count the number of hits recorded from each individual IP address that accessed it. Solution Many of the chores of administering a web server have to do with analyzing Apache logs, which Python makes easy: def calculateApacheIpHits(logfile_pathname): ''' return a dict mapping IP addresses to hit counts ''' ipHitListing = { } contents = open(logfile_pathname, "r") # go through each line of the logfile for line in contents: # split the string to isolate the IP address ip = line.split(" ", 1)[0] # Ensure length of the IP address is proper (see discussion) if 6 < len(ip) <= 15: # Increase by 1 if IP exists; else set hit count = 1 ipHitListing[ip] = ipHitListing.get(ip, 0) + 1 return ipHitListing Discussion This recipe supplies a function that returns a dictionary containing the hit counts for each individual IP address that has accessed your Apache web server, as recorded in an Apache log file. For example, a typical use would be: HitsDictionary = calculateApacheIpHits( "/usr/local/nusphere/apache/logs/access_log") This function has many quite useful applications. For example, I often use it in my code to determine the number of hits that are actually originating from locations other than my local host. This function is also used to chart which IP addresses are most actively viewing the pages that are served by a particular installation of Apache. This function performs a modest validation of each IP address, which is really just a length check: an IP address cannot be longer than 15 characters (4 sets of triplets and 3 periods) nor shorter than 7 (4 sets of single digits and 3 periods). This validation is not stringent, but it does reduce, at tiny runtime cost, the probability of placing into the dictionary some data that is obviously garbage. As a general technique, low-cost, highly approximate sanity checks for data that is expected to be OK (but one never knows for sure) are worth considering. However, if you want to be stricter, regular expressions can help. Change the loop in this recipe's function's body to: import re # an IP is: 4 strings, each of 1-3 digits, joined by periods ip_specs = r'\.'.join([r'\d{1,3}']*4) re_ip = re.compile(ip_specs) for line in contents: match = re_ip.match(line) if match: # Increase by 1 if IP exists; else set hit count = 1 ip = match.group( ) ipHitListing[ip] = ipHitListing.get(ip, 0) + 1 In this variant, we use a regular expression to extract and validate the IP at the same time. This approach enables us to avoid the split operation as well as the length check, and thus amortizes most of the runtime cost of matching the regular expression. This variant is only a few percentage points slower than the recipe's solution. Of course, the pattern given here as ip_specs is not entirely precise either, since it accepts, as components of an IP quad, arbitrary strings of one to three digits, while the components should be more constrained. But to ward off garbage lines, this level of sanity check is sufficient. Another alternative is to convert and check the address: extract string ip just as we do in this recipe's Solution, then: # Ensure the IP address is proper try: quad = map(int, ip.split('.')) except ValueError: pass else: if len(quad)==4 and min(quad)>=0 and max(quad)<=255: # Increase by 1 if IP exists; else set hit count = 1 ipHitListing[ip] = ipHitListing.get(ip, 0) + 1 This approach is more work, but it does guarantee that only IP addresses that are formally valid get counted at all. See Also The Apache web server is available and documented at http://httpd.apache.org; regular expressions are covered in the docs of the re module in the Library Reference and Python in a Nutshell. Recipe 10.5. Calculating the Rate of Client Cache Hits on Apache Credit: Mark Nenadov Problem You need to monitor how often client requests are refused by your Apache web server because the client's cache of the page is already up to date. Solution When a browser queries a server for a page that the browser has in its cache, the browser lets the server know about the cached data, and the server returns a special error code (rather than serving the page again) if the client's cache is up to date. Here's how to find the statistics for such occurrences in your server's logs: def clientCachePercentage(logfile_pathname): contents = open(logfile_pathname, "r") totalRequests = 0 cachedRequests = 0 for line in contents: totalRequests += 1 if line.split(" ")[8] == "304": # if server returned "not modified" cachedRequests += 1 return int(0.5+float(100*cachedRequests)/totalRequests) Discussion The percentage of requests to your Apache server that are met by the client's own cache is an important factor in the perceived performance of your server. The code in this recipe helps you get this information from the server's log. Typical use would be: log_path = "/usr/local/nusphere/apache/logs/access_log" print "Percentage of requests that were client-cached: " + str( clientCachePercentage(log_path)) + '%' The recipe reads the log file one line at a time by looping over the filethe normal way to read a file nowadays. Trying to read the whole log file in memory, by calling the readlines method on the file object, would be an unsuitable approach for very large files, which server log files can certainly be. That approach might not work at all, or might work but damage performance considerably by swamping your machine's virtual memory. Even when it works, readlines offers no advantage over the approach used in this recipe. The body of the for loop calls the split method on each line string, with a string of a single space as the argument, to split the line into a tuple of its space-separated fields. Then it uses indexing ([8]) to get the ninth such field. Apache puts the error code into the ninth field of each line in the log. Code "304" means "not modified" (i.e., the client's cache was already correctly updated). Wecount those cases in the cachedRequests variable and all lines in the log in the totalRequests variable, so that, in the end, we can return the percentage of cache hits. The expression we use in the return statement computes the percentage as a float number, then rounds it correctly to the closest int, because an integer result is most useful in practice. See Also The Apache web server is available and documented at http://httpd.apache.org. Recipe 10.6. Spawning an Editor from a Script Credit: Larry Price, Peter Cogolo Problem You want users to work with their favorite text-editing programs to edit text files, to provide input to your script. Solution Module tempfile lets you create temporary files, and module os has many tools to check the environment and to work with files and external programs, such as text editors. A couple of functions can wrap this functionality into an easy-to-use form: import sys, os, tempfile def what_editor( ): editor = os.getenv('VISUAL') or os.getenv('EDITOR') if not editor: if sys.platform == 'windows': editor = 'Notepad.Exe' else: editor = 'vi' return editor def edited_text(starting_text=''): temp_fd, temp_filename = tempfile.mkstemp(text=True) os.write(temp_fd, starting_text) os.close(temp_fd) editor = what_editor( ) x = os.spawnlp(os.P_WAIT, editor, editor, temp_filename) if x: raise RuntimeError, "Can't run %s %s (%s)" % (editor, temp_filename, x) result = open(temp_filename).read( ) os.unlink(temp_filename) return result if _ _name_ _=='_ _main_ _': text = edited_text('''Edit this text a little, go ahead, it's just a demonstration, after all...! ''') print 'Edited text is:', text Discussion Your scripts may often need a substantial amount of textual input from the user. Letting users edit the text with their favorite text editor is an excellent feature for your script to have, and this recipe shows how you can obtain it. I have used variants of this approach for such purposes as adjusting configuration files, writing blog posts, and sending emails. If your scripts do not need to run on Windows, a more secure and slightly simpler way to codefunction edited_text is available: def edited_text(starting_text=''): temp_file = tempfile.NamedTemporaryFile( ) temp_file.write(starting_text) temp_file.seek(0) editor = what_editor( ) x = os.spawnlp(os.P_WAIT, editor, editor, temp_file.name) if x: raise RuntimeError, "Can't run %s %s (%s)" % (editor, temp_file.name, x) return temp_file.read( ) Unfortunately, this alternative relies on the editor we're spawning being able to open and modify the temporary file while we are holding that file open, and this capability is not supported on most versions of Windows. The version of edited_text given in the recipe is more portable. When you're using this recipe to edit text files that must respect some kind of syntax or other constraints, such as a configuration file, you can make your script simpler and more effective by using a cycle of "input/parse/re-edit in case of errors," providing immediate feedback to users when you can diagnose they've made a mistake in editing the file. Ideally, in such cases, you should reopen the editor already pointing at the line in error, which is possible with most Unix editors by passing them a first argument such as '+23', specifying that they start editing at line 23, before the filename argument. Unfortunately, such an argument would confuse many Windows editors, so you have to make some hard decisions here (if you do need to support Windows). See Also Documentation for modules tempfile and os in the Library Reference and Python in a Nutshell. Recipe 10.7. Backing Up Files Credit: Anand Pillai, Tiago Henriques, Mario Ruggier Problem You want to make frequent backup copies of all files you have modified within a directory tree, so that further changes won't accidentally obliterate some of your editing. Solution Version-control systems, such as RCS, CVS, and SVN, are very powerful and useful, but sometimes a simple script that you can easily edit and customize can be even handier. The following script checks for new files to back up in a tree that you specify. Run the script periodically to keep your backup copies up to date. import sys, os, shutil, filecmp MAXVERSIONS=100 def backup(tree_top, bakdir_name='bakdir'): for dir, subdirs, files in os.walk(tree_top): # ensure each directory has a subdir called bakdir backup_dir = os.path.join(dir, bakdir_name) if not os.path.exists(newdir): os.makedirs(newdir) # stop any recursing into the backup directories subdirs[:] = [d for d in subdirs if d != bakdir_name] for file in files: filepath = os.path.join(dir, file) destpath = os.path.join(backup_dir, file) # check existence of previous versions for index in xrange(MAXVERSIONS): backup = '%s.%2.2d' % (destpath, index) if not os.path.exists(backup): break if index > 0: # no need to backup if file and last version are identical old_backup = '%s.%2.2d' % (destpath, index-1) try: if os.path.isfile(old_backup ) and filecmp.cmp(abspath, old_backup, shallow=False): continue except OSError: pass try: shutil.copy(filepath, backup) except OSError: pass if _ _name_ _ == '_ _main_ _': # run backup on the specified directory (default: the current directory) try: tree_top = sys.argv[1] except IndexError: tree_top = '.' backup(tree_top)Discussion Although version-control systems are more powerful, this script can be useful in development work. I often customize it, for example, to keep backups only of files with certain extensions (or, when that's handier, of all files except those with certain extensions); it suffices to add an appropriate test at the very start of the for file in files loop, such as: name, ext = os.path.splitext(file) if ext not in ('.py', '.txt', '.doc'): continue This snippet first uses function splitext from the standard library module os.path to extract the file extension (starting with a period) into local variable ext, then conditionally executes statement continue, which passes to the next leg of the loop, unless the extension is one of a few that happen to be the ones of interest in the current subtree. Other potentially useful variants include backing files up to some other subtree (potentially on a removable drive, which has some clear advantages for backup purposes) rather than the current one, compressing the files that are being backed up (look at standard library module gzip for this purpose), and more refined ones yet. However, rather than complicating function backup by offering all of these variants as options, I prefer to copy the entire script to the root of each of the various subtrees of interest, and customize it with a little simple editing. While this strategy would be a very bad one for any kind of complicated, highly reusable production-level code, it is reasonable for a simple, straightforward system administration utility such as the one in this recipe. Worthy of note in this recipe's implementation is the use of function os.walk, a generator from the standard Python library's module os, which makes it very simple to iterate over all or most of a filesystem subtree, with no need for such subtleties as recursion or callbacks, just a straightforward for statement. To avoid backing up the backups, this recipe uses one advanced feature of os.walk: the second one of the three values that os.walk yields at each step through the loop is a list of subdirectories of the current directory. We can modify this list in place, removing some of the subdirectory names it contains. When we perform such an in-place modification, os.walk does not recurse through the subdirectories whose names we removed. The following steps deal only with the subdirectories whose names are left in. This subtle but useful feature of os.walk is one good example of how a generator can receive information from the code that's iterating on it, to affect details of the iteration being performed. See Also Documentation of standard library modules os, shutils, and gzip in the Library Reference and Python in a Nutshell. Recipe 10.8. Selectively Copying a Mailbox File Credit: Noah Spurrier, Dave Benjamin Problem You need to selectively copy a large mailbox file (in mbox style), passing each message through a filtering function that may alter or skip the message. Solution The Python Standard Library package email is the modern Python approach for this kind of task. However, standard library modules mailbox and rfc822 can also supply the base functionality to implement this task: def process_mailbox(mailboxname_in, mailboxname_out, filter_function): mbin = mailbox.PortableUnixMailbox(file(mailboxname_in,'r')) fout = file(mailboxname_out, 'w') for msg in mbin: if msg is None: break document = filter_function(msg, msg.fp.read( )) if document: assert document.endswith('\n\n') fout.write(msg.unixfrom) fout.writelines(msg.headers) fout.write('\n') fout.write(document) fout.close( ) Discussion I often write lots of little scripts to filter my mailbox, so I wrote this recipe's small module. I can import the module from each script and call the module's function process_mailbox as needed. Python's future direction is to perform email processing with the standard library package email, but lower-level modules, such as mailbox and rfc822, are still available in the Python Standard Library. They are sometimes easier to use than the rich, powerful, and very general functionality offered by package email. The function you pass to process_mailbox as the third argument, filter_function, must take two argumentsmsg, an rfc822 message object, and document, a string that is the message's entire body, ending with two line-end characters (\n\n). filter_function can return False, meaning that this message must be skipped (i.e., not copied at all to the output), or else it must return a string terminated with \n\n that is written to the output as the message body. Normally, filter_function returns either False or the same document argument it was called with, but in some cases you may find it useful to write to the output file an altered version of the message's body rather than the original message body. Here is an example of a filter function that removes duplicate messages: import sets found_ids = sets.Set( )def no_duplicates(msg, document): msg_id = msg.getheader('Message-ID') if msg_id in found_ids: return False found_ids.add(msg_id) return document In Python 2.4, you could use the built-in set rather than sets.Set, but for a case as simple as this, it makes no real difference in performance (and the usage is exactly the same, anyway). See Also Documentation about modules mailbox and rfc822, and package email, in the Library Reference and Python in a Nutshell. Recipe 10.9. Building a Whitelist of Email Addresses From a Mailbox Credit: Noah Spurrier Problem To help you configure an antispam system, you want a list of email addresses, commonly known as a whitelist, that you can trust won't send you spam. The addresses to which you send email are undoubtedly good candidates for this whitelist. Solution Here is a script to output "To" addresses given a mailbox path: #!/usr/bin/env python """ Extract and print all 'To:' addresses from a mailbox """ import mailbox def main(mailbox_path): addresses = { } mb = mailbox.PortableUnixMailbox(file(mailbox_path)) for msg in mb: toaddr = msg.getaddr('To')[1] addresses[toaddr] = 1 addresses = addresses.keys( ) addresses.sort( ) for address in addresses: print address if _ _name_ _ == '_ _main_ _': import sys main(sys.argv[1]) Discussion In addition to bypassing spam filters, identifying addresses of people you've sent mail to may also help in other ways, such as flagging emails from them as higher priority, depending on your mail-reading habits and your mail reader's capabilities. As long as your mail reader keeps mail you have sent in some kind of "Sent Items" mailbox in standard mailbox format, you can call this script with the path to the mailbox as its only argument, and the addresses to which you've sent mail will be emitted to standard output. The script is simple because the Python Standard Library module mailbox does all the hard work. All the script needs to do is collect the set of email addresses as it loops through all messages, then emit them. While collecting, we keep addresses as a dictionary, since that's much faster than keeping a list and checking each toaddr in order to append it only if it wasn't already in the list. When we're done collecting, we just extract the addresses from the dictionary as a list because we want to emit its items in sorted order. In Python 2.4, function main can be made even slightly more elegant, thanks to the new built-ins set and sorted: def main(mailbox_path): addresses = set( ) mb = mailbox.PortableUnixMailbox(file(mailbox_path)) for msg in mb: toaddr = msg.getaddr('To')[1] addresses.add(toaddr) for address in sorted(addresses): print address If your mailbox is not in the Unix mailbox style supported by mailbox.PortableUnixMailbox, you may want to use other classes supplied by the Python Standard Library module mailbox. For example, if your mailbox is in Qmail maildir format, you can use the mailbox.Maildir class to read it. See Also Documentation of the standard library module mailbox in the Library Reference and Python in a Nutshell. Recipe 10.10. Blocking Duplicate Mails Credit: Marina Pianu, Peter Cogolo Problem Many of the mails you receive are duplicates. You need to block the duplicates with a fast, simple filter before they reach a more time-consuming step, such as an anti-spam filter, in your email pipeline. Solution Many mail systems, such as the popular procmail, and KDE's KMail, enable you to control your mail-reception pipeline. Specifically, you can insert in the pipeline your filter programs, which get messages on standard input, may modify them, and emit them again on standard output. Here is one such filter, with the specific purpose of performing the task described in the Problemblocking messages that are duplicates of other messages that you have received recently: #!/usr/bin/python import time, sys, os, email now = time.time( ) # get archive of previously-seen message-ids and times kde_dir = os.expanduser('~/.kde') if not os.path.isdir(kde_dir): os.mkdir(kde_dir) arfile = os.path.join(kde_dir, 'duplicate_mails') duplicates = { } try: archive = open(arfile) except IOError: pass else: for line in archive: when, msgid = line[:-1].split(' ', 1) duplicates[msgid] = float(when) archive.close( ) redo_archive = False # suck message in from stdin and study it msg = email.message_from_file(sys.stdin) msgid = msg['Message-ID'] if msgid: if msgid in duplicates: # duplicate message: alter its subject subject = msg['Subject'] if subject is None: msg['Subject'] = '**** DUP **** ' + msgid else: del msg['Subject'] msg['Subject'] = '**** DUP **** ' + subject else: # non-duplicate message: redo the archive file redo_archive = True duplicates[msgid] = now else: # invalid (missing message-id) message: alter its subject subject = msg['Subject'] if subject is None: msg['Subject'] = '**** NID **** ' else: del msg['Subject'] msg['Subject'] = '**** NID **** ' + subject # emit message back to stdout print msg if redo_archive: # redo archive file, keep only msgs from the last two hours keep_last = now - 2*60*60.0 archive = file(arfile, 'w') for msgid, when in duplicates.iteritems( ): if when > keep_last: archive.write('%9.2f %s\n' % (when, what)) archive.close( ) Discussion Whether it is because of spammers' malice or incompetence, or because of hiccups at my Internet ISP (Internet service provider), at times I get huge amounts of duplicate messages that can overload my mail-reception pipeline, particularly antispam filters. Fortunately, like many other mail systems, KDE's KMail, the one I use, lets me insert my own filters in the mail reception pipeline. In particular, I can diagnose duplicate messages, alter their headers (I use "Subject" for clarity), and tell later stages in the filters' pipeline to throw away messages with such subjects or to shunt them aside into a dedicated mailbox for later perusal, without passing them on to the antispam and other filters. The email module from the Python Standard Library performs all the required parsing of the message and lets me access headers with dictionary-like indexing syntax. I need some "memory" of recently seen messages. Fortunately, I have noticed all duplicates happen within a few minutes of each other, so I don't have to keep that memory for longtwo hours are plenty. Therefore, I keep that memory in a simple text file, which records the time when a message was received and the message ID. I thought I might have to find a more advanced way to keep this kind of FIFO (first-in, first-out) archive, but I tried a simple approach firsta simple text file that is entirely rewritten whenever a new nonduplicate message arrives. This approach appears to perform quite adequately for my needs (at most a couple hundred messages an hour), even on my somewhat dated PC. "Do the simplest thing that could possibly work" strikes again! See Also Documentation about package email and modules time, sys and os in the Library Reference and Python in a Nutshell. Recipe 10.11. Checking Your Windows Sound System Credit: Anand Pillai Problem You need to check whether the sound subsystem on your Windows PC is properly configured. Solution The winsound module of the Python Standard Library makes this check really simple: import winsound try: winsound.PlaySound("*", winsound.SND_ALIAS) except RuntimeError, e: print 'Sound system has problems,', e else: print 'Sound system is OK' Discussion The sound system might pass this test and still be unable to produce sound correctly, due to a variety of possible problemsstarting from simple ones such as powered loudspeakers being turned off (there's no sensible way you can check for that in your program!), all the way to extremely subtle and complicated ones. When sound is a problem in your applications, using this recipe at least you know whether you should be digging into a subtle issue of device driver configuration or start by checking whether the loudspeakers are on! See Also Documentation on the Python Standard Library winsound module. Recipe 10.12. Registering or Unregistering a DLL on Windows Credit: Bill Bell Problem You want to register or unregister a DLL in Windows, just as it is normally done by regsrv32.exe, but you want to do it from Python, without requiring that executable to be present or bothering to find it. Solution All that Microsoft's regsrv32.exe does is load a DLL and call its entries named DllRegisterServer or DllUnregisterServer. This behavior is very easy to replicate via Thomas Heller's ctypes extension: from ctypes import windll dll = windll[r'C:\Path\To\Some.DLL'] result = dll.DllRegisterServer( ) result = dll.DllUnregisterServer( ) The result is of Windows type HRESULT, so, if you wish, ctypes can also implicitly check it for you, raising a ctypes.WindowsError exception when an error occurs; you just need to use ctypes.oledll instead of ctypes.windll. In other words, to have the result automatically checked and an exception raised in case of errors, instead of the previous script, use this one: from ctypes import oledll dll = oledll[r'C:\Path\To\Some.DLL'] dll.DllRegisterServer( ) dll.DllUnregisterServer( ) Discussion Thomas Heller's ctypes enables your Python code to load DLLs on Windows (and similar dynamic/shared libraries on other platforms) and call functions from such libraries, and it manages to perform these tasks with a high degree of both power and elegance. On Windows, in particular, it offers even further "added value" through such mechanisms as the oledll object, which, besides loading DLLs and calling functions from them, also checks the returned hrESULT instances and raises appropriate exceptions when the HRESULT values indicate errors. In this recipe, we're using ctypes (either the windll or oledll objects from that module) specifically to avoid the need to use Microsoft's regsrv32.exe to register or unregister DLLs that implement in-process COM servers for some CLSIDs. (A CLSID is a globally unique identifier that identifies a COM class object, and the abbreviation presumably stands for class identifier.) The cases in which you'll use this specific recipe are only those in which you need to register or unregister such COM DLLs (whether they're implemented in Python or otherwise makes no difference). Be aware, however, that the applicability of ctypes is far wider, as it extends to any case in which you wish your Python code to load and interact with a DLL (or, on platforms otherthan Windows, equivalent dynamically loaded libraries, such as .so files on Linux and .dynlib files on Mac OS X). The protocol that regsrv32.exe implements is well documented and very simple, so our own code can reimplement it in a jiffy. That's much more practical than requiring regsrv32.exe to be installed on the machine on which we want to register or unregister the DLLs, not to mention finding where the EXE file might be to run it directly (via os.spawn or whatever) and also finding an effective way to detect errors and show them to the user. See Also ctypes is at http://sourceforge.net/projects/ctypes. Recipe 10.13. Checking and Modifying the Set of Tasks Windows Automatically Runs at Login Credit: Daniel Kinnaer Problem You need to check which tasks Windows is set to automatically run at login and possibly change this set of tasks. Solution When administering Windows machines, it's crucial to keep track of the tasks each machine runs at login. Like so many Windows tasks, this requires working with the registry, and standard Python module _winreg enables this: import _winreg as wr aReg = wr.ConnectRegistry(None, wr.HKEY_LOCAL_MACHINE) try: targ = r'SOFTWARE\Microsoft\Windows\CurrentVersion\Run' print "*** Reading from", targ, "***" aKey = wr.OpenKey(aReg, targ) try: for i in xrange(1024): try: n, v, t = wr.EnumValue(aKey, i) print i, n, v, t except EnvironmentError: print "You have", i, "tasks starting at logon" break finally: wr.CloseKey(aKey) print "*** Writing to", targ, "***" aKey = wr.OpenKey(aReg, targ, 0, wr.KEY_WRITE) try: try: wr.SetValueEx(aKey, "MyNewKey", 0, REG_SZ, r"c:\winnt\explorer.exe") except EnvironmentError: print "Encountered problems writing into the Registry..." raise finally: CloseKey(aKey) finally: CloseKey(aReg) Discussion The Windows registry holds a wealth of crucial system administration data, and the Python standard module _winreg makes it feasible to read and alter data held in the registry. One of theitems held in the Windows registry is a list of tasks to be run at login (in addition to other lists held elsewhere, such as the user-specific Startup folder that this recipe does not deal with). This recipe shows how to examine the registry list of login tasks, and how to add a task to the list so it is run at login. (This recipe assumes you have Explorer installed at the specific location c:\winnt. If you have it installed elsewhere, edit the recipe accordingly.) If you want to remove the specific key added by this recipe, you can use the following simple script: import _winreg as wr aReg = wr.ConnectRegistry(None, wr.HKEY_LOCAL_MACHINE) targ = r'SOFTWARE\Microsoft\Windows\CurrentVersion\Run' aKey = wr.OpenKey(aReg, targ, 0, wr.KEY_WRITE) wr.DeleteValue(aKey, "MyNewKey") wr.CloseKey(aKey) wr.CloseKey(aReg) The TRy/finally constructs used in the recipe are far more robust than the simple sequence of function calls used in this latest snippet, since they ensure that everything is closed correctly regardless of whether the intervening calls succeed or fail. This care and prudence are strongly advisable for scripts that are meant be run in production, particularly for system-administration scripts that must generally run with administrator privileges. Such scripts therefore might harm a system's setup if they don't clean up after themselves properly. However, you can omit the try/finally when you know the calls will succeed or don't care what happens if they fail. In this case, if you have successfully added a task with the recipe's script, the calls in this simple cleanup script should work just fine. See Also Documentation for the standard module _winreg in the Library Reference; Windows API documentation available from Microsoft (http://msdn.microsoft.com); information on what is where in the registry tends to be spread information among many sources, but for some useful collections of such information, see http://www.winguides.com/registry and http://www.activewin.com/tips/reg/index.shtml. Recipe 10.14. Creating a Share on Windows Credit: John Nielsen Problem You want to share a folder of your Windows PC on a LAN. Solution PyWin32's win32net module makes this task very easy: import win32net import win32netcon shinfo={ } shinfo['netname'] = 'python test' shinfo['type'] = win32netcon.STYPE_DISKTREE shinfo['remark'] = 'data files' shinfo['permissions'] = 0 shinfo['max_uses'] = -1 shinfo['current_uses'] = 0 shinfo['path'] = 'c:\\my_data' shinfo['passwd'] = '' server = 'servername' win32net.NetShareAdd(server, 2, shinfo) Discussion While the task of sharing a folder is indeed fairly easy to accomplish, finding the information on how you do so isn't. All I could find in the win32net documentation was that you needed to pass a dictionary holding the share's data "in the format of SHARE_INFO_*." I finally managed to integrate this tidbit with the details from the Windows SDK (http://msdn.microsoft.com) and produce the information in this recipe. One detail that took me some effort to discover is that the constants you need to use as the value for the 'type' enTRy are "hidden away" in the win32netcon module. See Also PyWin32 docs at http://sourceforge.net/projects/pywin32/; Microsoft's MSDN site, http://msdn.microsoft.com. Recipe 10.15. Connecting to an Already Running Instance of Internet Explorer Credit: Bill Bell, Graham Fawcett Problem Instantiating Internet Explorer to access its interfaces via COM is easy, but you want to connect to an already running instance. Solution The simplest approach is to rely on Internet Explorer's CLSID: from win32com.client import Dispatch ShellWindowsCLSID = '{9BA05972-F6A8-11CF-A442-00A0C90A8F39}' ShellWindows = Dispatch(ShellWindowsCLSID) print '%d instances of IE' % len(shellwindows) print for shellwindow in ShellWindows : print shellwindow print shellwindos.LocationName print shellwindos.LocationURL print Discussion Dispatching on the CLSID provides a sequence of all the running instances of the application with that class. Of course, there could be none, one, or more. If you're interested in a specific instance, you may be able to identify it by checking, for example, for its properties LocationName and LocationURL. You'll see that Windows Explorer and Internet Explorer have the same CLSIDthey're basically the same application. If you need to distinguish between them, you can try adding at the start of your script the statement: from win32gui import GetClassName and then checking each shellwindow in the loop with: if GetClassName(shellwindow.HWND) == 'IEFrame': ... 'IEFrame' is supposed to result from this call (according to the docs) for all Internet Explorer instances and those only. However, I have not found this check to be wholly reliable across all versions and patch levels of Windows and Internet Explorer, so, take this approach as just one possibility (which is why I haven't added this check to the recipe's official "Solution").This recipe does not let you receive IE events. The most important event is probably DocumentComplete. You can roughly substitute checks on the Busy property for the inability to wait for that event, but remember not to poll too frequently (for that or any other property) or you may slow down your PC excessively. Something like: while shellwindow.Busy: time.sleep(0.2) Sleeping 0.2 seconds between checks may be a reasonable compromise between responding promptly and not loading your PC too heavily with a busy-waiting-loop. See Also PyWin32 docs at http://sourceforge.net/projects/pywin32/; Microsoft's MSDN site, http://msdn.microsoft.com. Recipe 10.16. Reading Microsoft Outlook Contacts Credit: Kevin Altis Problem Your Microsoft Outlook Contacts house a wealth of useful information, and you need to extract some of it in text form. Solution Like many other problems of system administration on Windows, this one is best approached by using COM. The most popular way to interface Python to COM is to use the win32com package, which is part of Mark Hammond's pywin32 extension package: from win32com.client import gencache, constants DEBUG = False class MSOutlook(object): def _ _init_ _(self): try: self.oOutlookApp = gencache.EnsureDispatch("Outlook.Application") self.outlookFound = True except: print "MSOutlook: unable to load Outlook" self.outlookFound = False self.records = [ ] def loadContacts(self, keys=None): if not self.outlookFound: return onMAPI = self.oOutlookApp.GetNamespace("MAPI") ofContacts = onMAPI.GetDefaultFolder(constants.olFolderContacts) if DEBUG: print "number of contacts:", len(ofContacts.Items) for oc in range(len(ofContacts.Items)): contact = ofContacts.Items.Item(oc + 1) if contact.Class == constants.olContact: if keys is None: # no keys were specified, so build up a list of all keys # that belong to some types we know we can deal with good_types = int, str, unicode keys = [key for key in contact._prop_map_get_ if isinstance(getattr(contact, key), good_types) ] if DEBUG: print "Fields\n== == == == == == == == == == == ==" keys.sort( ) for key in keys: print key record = { } for key in keys: record[key] = getattr(contact, key) self.records.append(record) if DEBUG: print oc, contact.FullName if _ _name_ _ == '_ _main_ _': if '-d' in sys.argv: DEBUG = True if DEBUG: print "attempting to load Outlook" oOutlook = MSOutlook( ) if not oOutlook.outlookFound: print "Outlook not found" sys.exit(1) fields = ['FullName', 'CompanyName', 'MailingAddressStreet', 'MailingAddressCity', 'MailingAddressState', 'MailingAddressPostalCode', 'HomeTelephoneNumber', 'BusinessTelephoneNumber', 'MobileTelephoneNumber', 'Email1Address', 'Body', ] if DEBUG: import time print "loading records..." startTime = time.time( ) # to get all fields just call oOutlook.loadContacts( ) # but getting a specific set of fields is much faster oOutlook.loadContacts(fields) if DEBUG: print "loading took %f seconds" % (time.time( ) - startTime) print "Number of contacts: %d" % len(oOutlook.records) print "Contact: %s" % oOutlook.records[0]['FullName'] print "Body:\n%s" % oOutlook.records[0]['Body'] Discussion This recipe's code could use more error-checking, and you could get it by using nested TRy/except blocks, but I didn't want to obscure the code's fundamental simplicity in this recipe. This recipe should work with different versions of Outlook, but I've tested it only with Outlook 2000. If you have applied the Outlook security patches then you will be prompted with a dialog requesting access to Outlook for 1-10 minutes from an external program, which in this case is Python. The code has already been optimized in two important ways. First, by ensuring that the Python COM wrappers for Outlook have been generated, which is guaranteed by calling gencache.EnsureDispatch. Second, in the loop that reads the contacts, the Contact reference is obtained only once and then kept in a local variable contact to avoid repeated references. This simple but crucial optimization is the role of the statement: contact = ofContacts.Items.Item(oc + 1) Both of these optimizations have a dramatic impact on total import time, and both are important enough to keep in mind. Specifically, the EnsureDispatch idea is important for most uses of COM in Python; the concept of getting an object reference, once, into a local variable (rather than repeating indexing, calls, and attribute accesses) is even more important and applies to every use of Python. Simple variations of this script can be applied to other elements of the Outlook object model such as the Calendar and Tasks. You'll want to look at the Python wrappers generated for Outlook in the C:\Python23\Lib\site-packages\win32com\gen_py directory. I also suggest that you look at the Outlook object model documentation on MSDN and/or pick up a book on the subject. See AlsoPyWin32 docs at http://sourceforge.net/projects/pywin32/; Microsoft's MSDN site, http://msdn.microsoft.com. Recipe 10.17. Gathering Detailed System Informationon Mac OS X Credit: Brian Quinlan Problem You want to retrieve detailed information about a Mac OS X system. You want either complete information about the system or information about particular keys in the system-information database. Solution Mac OS X's system_profiler command can provide system information as an XML stream that we can parse and examine: #!/usr/bin/env python from xml import dom from xml.dom.xmlbuilder import DOMInputSource, DOMBuilder import datetime, time, os def group(seq, n): """group([0, 3, 4, 10, 2, 3, 1], 3) => [(0, 3, 4), (10, 2, 3)] Group a sequence into n-subseqs, discarding incomplete subseqs. """ return [ seq[i:i+n] for i in xrange(0, len(seq)-n+1, n) ] def remove_whitespace_nodes(node): """Removes all of the whitespace-only text descendants of a DOM node.""" remove_list = [ ] for child in node.childNodes: if child.nodeType == dom.Node.TEXT_NODE and not child.data.strip( ): remove_list.append(child) elif child.hasChildNodes( ): remove_whitespace_nodes(child) for child in remove_list: node.removeChild(child) child.unlink( ) class POpenInputSource(DOMInputSource): "Use stdout from an external program as a DOMInputSource" def _ _init_ _(self, command): super(DOMInputSource, self)._ _init_ _( ) self.byteStream = os.popen(command) class OSXSystemProfiler(object): "Provide information from the Mac OS X System Profiler" def _ _init_ _(self, detail=-1): """detail can range from -2 to +1. Larger numbers return more info. Beware of +1, can take many minutes to get all info!""" b = DOMBuilder( ) self.document = b.parse( POpenInputSource('system_profiler -xml -detailLevel %d' % detail)) remove_whitespace_nodes(self.document) def _content(self, node): "Get the text node content of an element, or an empty string" if node.firstChild: return node.firstChild.nodeValue else: return '' def _convert_value_node(self, node): """Convert a 'value' node (i.e. anything but 'key') into a Python data structure""" if node.tagName == 'string': return self._content(node) elif node.tagName == 'integer': return int(self._content(node)) elif node.tagName == 'real': return float(self._content(node)) elif node.tagName == 'date': # 2004-07-05T13:29:29Z return datetime.datetime( *time.strptime(self._content(node), '%Y-%m-%dT%H:%M:%SZ')[:5]) elif node.tagName == 'array': return [self._convert_value_node(n) for n in node.childNodes] elif node.tagName == 'dict': return dict([(self._content(n), self._convert_value_node(m)) for n, m in group(node.childNodes, 2)]) else: raise ValueError, 'Unknown tag %r' % node.tagName def _ _getitem_ _(self, key): from xml import xpath # pyxml's xpath does not support /element1[...]/element2... nodes = xpath.Evaluate('//dict[key=%r]' % key, self.document) results = [ ] for node in nodes: v = self._convert_value_node(node)[key] if isinstance(v, dict) and '_order' in v: # this is just information for display pass else: results.append(v) return results def all(self): """Return the complete information from the system profiler as a Python data structure""" return self._convert_value_node( self.document.documentElement.firstChild) def main( ): from optparse import OptionParser from pprint import pprint info = OSXSystemProfiler( ) parser = OptionParser( ) parser.add_option("-f", "--field", action="store", dest="field", help="display the value of the specified field") options, args = parser.parse_args( ) if args: parser.error("no arguments are allowed") if options.field is not None: pprint(info[options.field]) else: # print some keys known to exist in only one important dict for k in ['cpu_type', 'current_processor_speed', 'l2_cache_size', 'physical_memory', 'user_name', 'os_version', 'ip_address']: print '%s: %s' % (k, info[k][0]) if _ _name_ _ == '_ _main_ _': main( )Discussion Mac OS X puts at your disposal a wealth of information about your system through the system_profiler application. This recipe shows how to access that information from your Python code. First, you have to instantiate class OSXSystemProfiler, for example, via a statement such as info = OSXSystemProfiler( ); once you have done that, you can obtain all available information by calling info.all( ), or information for one specific key by indexing info[thekey]. The main function in the recipe, which executes when you run this module as a main script, emits information to standard outputeither a specific key, requested by using switch -f when invoking the script, or, by default, a small set of keys known to be generally useful. For example, when run on the old Apple iBook belonging to one of this book's editors (no prize for guessing which one), the script in this recipe emits the following output: cpu_type: PowerPC G4 (3.3) current_processor_speed: 800 MHz l2_cache_size: 256 KB physical_memory: 640 MB user_name: Alex (alex) os_version: Mac OS X 10.3.6 (7R28) ip_address: [u'192.168.0.190'] system_profiler returns XML data in pinfo format, so this recipe implements a partial pinfo parser, using Python's standard library XML-parsing facilities, and the xpath implementation from the PyXML extensions. More information about Python's facilities that help you deal with XML can be found in Chapter 12. See Also Documentation of the standard Python library support for XML in the Library Reference and Python in a Nutshell; PyXML docs at http://pyxml.sourceforge.net/; Mac OS X system_profiler docs at http://developer.apple.com/documentation/Darwin/Reference/ManPages/man8/system_profiler.8.html; Chapter 12. Chapter 11. User Interfaces Introduction Recipe 11.1. Showing a Progress Indicator on a Text Console Recipe 11.2. Avoiding lambda in Writing Callback Functions Recipe 11.3. Using Default Values and Bounds with tkSimpleDialog Functions Recipe 11.4. Adding Drag and Drop Reordering to a Tkinter Listbox Recipe 11.5. Entering Accented Characters in Tkinter Widgets Recipe 11.6. Embedding Inline GIFs Using Tkinter Recipe 11.7. Converting Among Image Formats Recipe 11.8. Implementing a Stopwatch in Tkinter Recipe 11.9. Combining GUIs and Asynchronous I/Owith Threads Recipe 11.10. Using IDLE's Tree Widget in Tkinter Recipe 11.11. Supporting Multiple Values per Row in a Tkinter Listbox Recipe 11.12. Copying Geometry Methods and Options Between Tkinter Widgets Recipe 11.13. Implementing a Tabbed Notebook for Tkinter Recipe 11.14. Using a wxPython Notebook with Panels Recipe 11.15. Implementing an ImageJ Plug-in in Jython Recipe 11.16. Viewing an Image from a URL with Swing and Jython Recipe 11.17. Getting User Input on Mac OS Recipe 11.18. Building a Python Cocoa GUI Programmatically Recipe 11.19. Implementing Fade-in Windows with IronPython Introduction Credit: Fredrik Lundh, SecretLabs AB, author of Python Standard Library Back in the early days of interactive computing, most computers offered terminals that looked and behaved pretty much like clunky typewriters. The main difference from an ordinary typewriter was that the computer was in the loop. It could read what the user typed and print hard-copy output on a roll of paper. So when you found yourself in front of a 1960s Teletype ASR-33, the only reasonable way to communicate with the computer was to type a line of text, press the send key, hope that the computer would manage to figure out what you meant, and wait for the response to appear on the paper roll. This line-oriented way of communicating with your computer is known as a command-line interface (CLI). Some 40 years later, the paper roll has been replaced with high-resolution video displays, which can display text in multiple typefaces, color photographs, and even animated 3D graphics. The keyboard is still around, but we also have pointing devices such as the mouse, trackballs, game controls, touchpads, and other input devices. The combination of a graphics display and the mouse made it possible to create a new kind of user interface: the graphical user interface (GUI). When done right, a GUI can give the user a better overview of what a program can do (and what it is doing), and make it easier to carry out many kinds of tasks. However, most programming languages, including Python, make it easy to write programs using teletype-style output and input. In Python, you use the print statement to print text to the display and the input and raw_input functions to read expressions and text strings from the keyboard. Creating GUIs takes more work. You need access to functions to draw text and graphics on the screen, select typefaces and styles, and read information from the keyboard and other input devices. You need to write code to interact with other applications (via a window manager), keep your windows updated when the user moves them around, and respond to key presses and mouse actions. To make this a bit easier, programmers have developed graphical user interface toolkits, which provide standard solutions to these problems. A typical GUI toolkit provides a number of ready- made GUI building blocks, usually called widgets. Common standard widgets include text and image labels, buttons, and text-entry fields. Many toolkits also provide more advanced widgets, such as Tkinter's Text widget, which is a rather competent text editor/display component. All major toolkits are event based, which means that your program hands control over to the toolkit (usually by calling a "main loop" function or method). The toolkit then calls back into your application when certain events occurfor example, when the user clicks OK in a dialog or when a window needs to be redrawn. Most toolkits also provide ways to position widgets on the screen automatically (e.g., in tables, rows, or columns) and to modify widget behavior and appearance. Tkinter is the de facto standard toolkit for Python and comes with most Python distributions. Tkinter provides an object-oriented layer on top of the Tcl/Tk GUI library and runs on Windows, Unix, and Macintosh systems. Tkinter is easy to use but provides a relatively small number of standard widgets. Tkinter extension libraries, such as Pmw and Tix, supply many components missing from plain Tkinter, and you can use Tkinter's advanced Text and Canvas widgets to create custom widgets. The Widget Construction Kit, WCK, lets you write all sorts of new widgets in pure Python: see http://effbot.org/zone/wck.htm. wxPython (http://www.wxPython.org) is another popular toolkit; it is based on the wxWidgetsC++ library (http://www.wxWidgets.org). wxPython is modeled somewhat after the Windows MFC library but is available for multiple platforms. wxPython provides a rich set of widgets, and it's relatively easy to create custom widgets. PyGTK (http://www.pygtk.org) is an object-oriented Python interface to the Gimp toolkit (GTK), used in projects such as Gnome and the Gimp. PyGTK is a good choice for Linux applications, especially if you want them to run in the Gnome environment. PyQt (http://www.riverbankcomputing.co.uk/pyqt/index.php) is a Python wrapper for TrollTech's Qt library (http://www.trolltech.com), which is the basis of the popular KDE environment, as well as the Qtopia environment for hand-held computers; it also runs on Windows and Mac OS X. Qt and PyQt require license fees for commercial (software that is not free) use, but are free (licensed by the GPL) for free software development. (No GPL-licensed Qt is currently available for Windows, but one is under developmentsee http://kde-cygwin.sourceforge.net/qt3-win32/.) You can also use many other toolkits from Python. Mark Hammond's Pythonwin gives access to Windows MFC. Greg Ewing is developing a cross-platform GUI API, known as PyGUI (http://nz.cosc.canterbury.ac.nz/~greg/python_gui/), developed specifically for Python and taking advantage of Python's unique strengths. Also available are interfaces to Motif/X11 and Mac OS X native toolboxes and many other toolkits. Cameron Laird maintains a list of toolkits at http://starbase.neosoft.com/~claird/comp.lang.python/python_GUI.html. It currently lists about 20 toolkits. A Wiki page at http://www.python.org/cgi-bin/moinmoin/GuiProgramming is actively maintained lists even more. Finally, several projects, in various stages, are based on the idea of overlaying easy unified APIs on top of one or more other toolkits or graphical facilities. anygui (rather dormantsee http://www.anygui.org), PythonCard (pretty activesee http://pythoncard.sourceforge.net/), Wax (http://zephyrfalcon.org/labs/dope_on_wax.html), and PyUI (http://pyui.sourceforge.net/) are examples of this "higher-level" approach. Recipe 11.1. Showing a Progress Indicator on a Text Console Credit: Larry Bates Problem Your program has no GUI (i.e., your program just runs on a text console), and yet you want your program to show to the user a "progress indicator bar" during lengthy operations, to communicate that work is progressing and the amount of the total work that has been completed. Solution We can easily code a simple little class to handle this whole task: import sys class progressbar(object): def _ _init_ _(self, finalcount, block_char='.'): self.finalcount = finalcount self.blockcount = 0 self.block = block_char self.f = sys.stdout if not self.finalcount: return self.f.write('\n------------------ % Progress -------------------1\n') self.f.write(' 1 2 3 4 5 6 7 8 9 0\n') self.f.write('----0----0----0----0----0----0----0----0----0----0\n') def progress(self, count): count = min(count, self.finalcount) if self.finalcount: percentcomplete = int(round(100.0*count/self.finalcount)) if percentcomplete < 1: percentcomplete = 1 else: percentcomplete=100 blockcount = int(percentcomplete//2) if blockcount <= self.blockcount: return for i in range(self.blockcount, blockcount): self.f.write(self.block) self.f.flush( ) self.blockcount = blockcount if percentcomplete == 100: self.f.write("\n") Discussion Here is an example of the use of this progressbar class, presented, as usual, with a guard of if _ _name_ _ == '_ _main_ _'. We can make it part of the module containing the class and have it run when the module is executed as a "main script":if _ _name_ _ == "_ _main_ _": from time import sleep pb = progressbar(8, "*") for count in range(1, 9): pb.progress(count) sleep(0.2) pb = progressbar(100) pb.progress(20) sleep(0.3) pb.progress(47) sleep(0.3) pb.progress(90) sleep(0.3) pb.progress(100) print "testing 1:" pb = progressbar(1) pb.progress(1) Programs that run lengthy operations, such as FTP downloads and database insertions, should normally give visual feedback to the user regarding the progress of the task that is running. GUI toolkits generally have such facilities included as "widgets", but if your program does not otherwise require a GUI, it's overkill to give it one just to be able to display a progress bar. This recipe's progress bar class provides an easy way of showing the percentage of completion that is updated periodically by the program. The recipe operates on the basis of a totally arbitrary final count that the ongoing task is supposed to reach at the end. This makes it optimally easy for the application that makes use of the progressbar class: the application can use any handy unit of measure (such as amount of bytes downloaded for an FTP download, number of records inserted for a database insertion, etc.) to track the task's progress and completion. As long as the same unit of measure applies to both the "final count" and the count argument that the application must periodically pass to the progress method, the progress bar's display will be accurate. See Also Documentation on text-mode console I/O in Python in a Nutshell. Recipe 11.2. Avoiding lambda in Writing Callback Functions Credit: Danny Yoo, Martin Sjogren Problem You need to use many callbacks without arguments, typically while writing a Tkinter-based GUI, and you'd rather avoid using lambda. Solution Between the classic lambda approach and a powerful general-purpose currying mechanism is a third, extremely simple way for doing callbacks that can come in handy in many practical cases: def command(callback, *args, **kwargs): def do_call( ): return callback(*args, **kwargs) # 2.4 only: do_call._ _name_ _ = callback._ _name_ _ return do_call Discussion I remember a utility class (to perform the same task handled by a closure in this recipe) quite a while back, but I don't remember who to attribute it to. Perhaps I saw it in John E. Grayson, Python and Tkinter Programming (Manning). Writing a lot of callbacks that give customized arguments can look a little awkward with lambda, so this command closure provides alternative syntax that is easier to read. For example: import Tkinter def hello(name): print "Hello", name root = Tk( ) # the lambda way of doing it: Button(root, text="Guido", command=lambda name="Guido": hello(name)).pack( ) # using the Command class: Button(root, text="Guido", command=command(hello, "Guido")).pack( ) Of course, you can also use a more general currying approach, which enables you to fix some of the arguments when you bind the callback, while others may be given at call time (see Recipe 16.4). However, "doing the simplest thing that can possibly work" is a good programming principle (this wording of the principle is due, I believe, to Kent Beck). If your application needs callbacks that fix all arguments at currying time and others that leave some arguments to be determined at callback time, it's probably simpler to use the more general currying approach for all the callbacks. But if all the callbacks you need must fix all arguments at currying time, it may be simpler to forego unneeded generality and use the simpler, less-general approach in this recipe exclusively. You can always refactor later if it turns out that you do need the generality.See Also Recipe 16.4; information about Tkinter can be obtained from a variety of sources, such as Fredrik Lundh, An Introduction to Tkinter (PythonWare: http://www.pythonware.com/library), New Mexico Tech's Tkinter Reference (http://www.nmt.edu/tcc/help/lang/python/docs.html), Python in a Nutshell, and various other books. Recipe 11.3. Using Default Values and Bounds with tkSimpleDialog Functions Credit: Mike Foord, Peter Cogolo Problem You need to get an input value from the user with one of Tkinter's tkSimpleDialog dialog functions, but you want to add a default value, or to ensure that the value entered lies within certain bounds. Solution Each of Tkinter's tkSimpleDialog functions (askstring, askfloat, askinteger) supports an optional default value, as well as optional validation against minimum and maximum value. However, this set of features is not clearly spelled out in the documentation. Here's a wrapper function that you may find preferable: import tkSimpleDialog _dispatch = { str: tkSimpleDialog.askstring, int: tkSimpleDialog.askinteger, float: tkSimpleDialog.askfloat, } def getinput(title, prompt, type=str, default=None, min=None, max=None): ''' gets from the user an input of type `type' (str, int or float), optionally with a default value, and optionally constrained to lie between the values `min' and `max' (included). ''' f = _dispatch.get(type) if not f: raise TypeError, "Can't ask for %r input" % (type,) return f(title, prompt, initialvalue=default, minvalue=min, maxvalue=max) Discussion The built-in tkSimpleDialog module offers a few simple functions that pop up dialogs that ask the user to input a string, a float, or an integernot a very advanced user interface but dirt-simple to use in your programs. Unfortunately, while these functions do support a few nice extras (the ability to pass in a default value, and having the result validated within certain optional minimum and maximum values), the module's documentation (what little there is of it) does not make this feature clear. Even the pydoc-generated page http://epydoc.sourceforge.net/stdlib/public/tkSimpleDialog-module.html just says "see SimpleDialog class." Since no such class exists, seeing it is not easy. (The relevant class is actually named _QueryDialog, and due to the leading underscore in the name, it is considered "private". Therefore pydoc does not build a documentation web page for it.) This recipe shows how to access this functionality that's already part of the Python Standard Library. As a side benefit, it refactors the functionality into a single getinput function that takes as an argument the type of input desired (defaulting t