What’s New in Python 2.6

Python 3.0

The development cycle for Python versions 2.6 and 3.0 was synchronized, with the alpha and beta releases for both versions being made on the same days. The development of 3.0 has influenced many features in 2.6.

Python 3.0 is a far-ranging redesign of Python that breaks compatibility with the 2.x series. This means that existing Python code will need some conversion in order to run on Python 3.0. However, not all the changes in 3.0 necessarily break compatibility. In cases where new features won’t cause existing code to break, they’ve been backported to 2.6 and are described in this document in the appropriate place. Some of the 3.0-derived features are:

• A __complex__() method for converting objects to a complex number.

• Alternate syntax for catching exceptions: except TypeError as exc.

• The addition of functools.reduce() as a synonym for the built-in reduce() function.

Python 3.0 adds several new built-in functions and changes the semantics of some existing builtins. Functions that are new in 3.0 such as bin() have simply been added to Python 2.6, but existing builtins haven’t been changed; instead, the future_builtins module has versions with the new 3.0 semantics. Code written to be compatible with 3.0 can do from future_builtins import hex, map as necessary.

A new command-line switch, -3, enables warnings about features that will be removed in Python 3.0. You can run code with this switch to see how much work will be necessary to port code to 3.0. The value of this switch is available to Python code as the boolean variable sys.py3kwarning, and to C extension code as Py_Py3kWarningFlag.

Changes to the Development Process

While 2.6 was being developed, the Python development process underwent two significant changes: we switched from SourceForge’s issue tracker to a customized Roundup installation, and the documentation was converted from LaTeX to reStructuredText.

PEP 343: The ‘with’ statement

The previous version, Python 2.5, added the ‘with’ statement as an optional feature, to be enabled by a from __future__ import with_statement directive. In 2.6 the statement no longer needs to be specially enabled; this means that with is now always a keyword. The rest of this section is a copy of the corresponding section from the “What’s New in Python 2.5” document; if you’re familiar with the ‘with’ statement from Python 2.5, you can skip this section.

The ‘with’ statement clarifies code that previously would use try...finally blocks to ensure that clean-up code is executed. In this section, I’ll discuss the statement as it will commonly be used. In the next section, I’ll examine the implementation details and show how to write objects for use with this statement.

The ‘with’ statement is a control-flow structure whose basic structure is:

with expression [as variable]:
    with-block

The expression is evaluated, and it should result in an object that supports the context management protocol (that is, has __enter__() and __exit__() methods).

The object’s __enter__() is called before with-block is executed and therefore can run set-up code. It also may return a value that is bound to the name variable, if given. (Note carefully that variable is not assigned the result of expression.)

After execution of the with-block is finished, the object’s __exit__() method is called, even if the block raised an exception, and can therefore run clean-up code.

Some standard Python objects now support the context management protocol and can be used with the ‘with’ statement. File objects are one example:

with open('/etc/passwd', 'r') as f:
    for line in f:
        print line
        ... more processing code ...

After this statement has executed, the file object in f will have been automatically closed, even if the for loop raised an exception part-way through the block.

The threading module’s locks and condition variables also support the ‘with’ statement:

lock = threading.Lock()
with lock:
    # Critical section of code
    ...

The lock is acquired before the block is executed and always released once the block is complete.

The localcontext() function in the decimal module makes it easy to save and restore the current decimal context, which encapsulates the desired precision and rounding characteristics for computations:

from decimal import Decimal, Context, localcontext

# Displays with default precision of 28 digits
v = Decimal('578')
print v.sqrt()

with localcontext(Context(prec=16)):
    # All code in this block uses a precision of 16 digits.
    # The original context is restored on exiting the block.
    print v.sqrt()

db_connection = DatabaseConnection()
with db_connection as cursor:
    cursor.execute('insert into ...')
    cursor.execute('delete from ...')
    # ... more operations ...

class DatabaseConnection:
    # Database interface
    def cursor(self):
        "Returns a cursor object and starts a new transaction"
    def commit(self):
        "Commits current transaction"
    def rollback(self):
        "Rolls back current transaction"

class DatabaseConnection:
    ...
    def __enter__(self):
        # Code to start a new transaction
        cursor = self.cursor()
        return cursor

class DatabaseConnection:
    ...
    def __exit__(self, type, value, tb):
        if tb is None:
            # No exception, so commit
            self.commit()
        else:
            # Exception occurred, so rollback.
            self.rollback()
            # return False

from contextlib import contextmanager

@contextmanager
def db_transaction(connection):
    cursor = connection.cursor()
    try:
        yield cursor
    except:
        connection.rollback()
        raise
    else:
        connection.commit()

db = DatabaseConnection()
with db_transaction(db) as cursor:
    ...

lock = threading.Lock()
with nested (db_transaction(db), lock) as (cursor, locked):
    ...

import urllib, sys
from contextlib import closing

with closing(urllib.urlopen('http://www.yahoo.com')) as f:
    for line in f:
        sys.stdout.write(line)

PEP 366: Explicit Relative Imports From a Main Module

Python’s -m switch allows running a module as a script. When you ran a module that was located inside a package, relative imports didn’t work correctly.

The fix for Python 2.6 adds a __package__ attribute to modules. When this attribute is present, relative imports will be relative to the value of this attribute instead of the __name__ attribute.

PEP 302-style importers can then set __package__ as necessary. The runpy module that implements the -m switch now does this, so relative imports will now work correctly in scripts running from inside a package.

PEP 370: Per-user site-packages Directory

When you run Python, the module search path sys.path usually includes a directory whose path ends in "site-packages". This directory is intended to hold locally-installed packages available to all users using a machine or a particular site installation.

Python 2.6 introduces a convention for user-specific site directories. The directory varies depending on the platform:

• Unix and Mac OS X: ~/.local/

• Windows: %APPDATA%/Python

Within this directory, there will be version-specific subdirectories, such as lib/python2.6/site-packages on Unix/Mac OS and Python26/site-packages on Windows.

If you don’t like the default directory, it can be overridden by an environment variable. PYTHONUSERBASE sets the root directory used for all Python versions supporting this feature. On Windows, the directory for application-specific data can be changed by setting the APPDATA environment variable. You can also modify the site.py file for your Python installation.

The feature can be disabled entirely by running Python with the -s option or setting the PYTHONNOUSERSITE environment variable.

PEP 371: The multiprocessing Package

The new multiprocessing package lets Python programs create new processes that will perform a computation and return a result to the parent. The parent and child processes can communicate using queues and pipes, synchronize their operations using locks and semaphores, and can share simple arrays of data.

The multiprocessing module started out as an exact emulation of the threading module using processes instead of threads. That goal was discarded along the path to Python 2.6, but the general approach of the module is still similar. The fundamental class is the Process, which is passed a callable object and a collection of arguments. The start() method sets the callable running in a subprocess, after which you can call the is_alive() method to check whether the subprocess is still running and the join() method to wait for the process to exit.

Here’s a simple example where the subprocess will calculate a factorial. The function doing the calculation is written strangely so that it takes significantly longer when the input argument is a multiple of 4.

import time
from multiprocessing import Process, Queue

def factorial(queue, N):
    "Compute a factorial."
    # If N is a multiple of 4, this function will take much longer.
    if (N % 4) == 0:
        time.sleep(.05 * N/4)

    # Calculate the result
    fact = 1L
    for i in range(1, N+1):
        fact = fact * i

    # Put the result on the queue
    queue.put(fact)

if __name__ == '__main__':
    queue = Queue()

    N = 5

    p = Process(target=factorial, args=(queue, N))
    p.start()
    p.join()

    result = queue.get()
    print 'Factorial', N, '=', result

A Queue is used to communicate the result of the factorial. The Queue object is stored in a global variable. The child process will use the value of the variable when the child was created; because it’s a Queue, parent and child can use the object to communicate. (If the parent were to change the value of the global variable, the child’s value would be unaffected, and vice versa.)

Two other classes, Pool and Manager, provide higher-level interfaces. Pool will create a fixed number of worker processes, and requests can then be distributed to the workers by calling apply() or apply_async() to add a single request, and map() or map_async() to add a number of requests. The following code uses a Pool to spread requests across 5 worker processes and retrieve a list of results:

from multiprocessing import Pool

def factorial(N, dictionary):
    "Compute a factorial."
    ...
p = Pool(5)
result = p.map(factorial, range(1, 1000, 10))
for v in result:
    print v

This produces the following output:

1
39916800
51090942171709440000
8222838654177922817725562880000000
33452526613163807108170062053440751665152000000000
...

The other high-level interface, the Manager class, creates a separate server process that can hold master copies of Python data structures. Other processes can then access and modify these data structures using proxy objects. The following example creates a shared dictionary by calling the dict() method; the worker processes then insert values into the dictionary. (Locking is not done for you automatically, which doesn’t matter in this example. Manager’s methods also include Lock(), RLock(), and Semaphore() to create shared locks.)

import time
from multiprocessing import Pool, Manager

def factorial(N, dictionary):
    "Compute a factorial."
    # Calculate the result
    fact = 1L
    for i in range(1, N+1):
        fact = fact * i

    # Store result in dictionary
    dictionary[N] = fact

if __name__ == '__main__':
    p = Pool(5)
    mgr = Manager()
    d = mgr.dict()         # Create shared dictionary

    # Run tasks using the pool
    for N in range(1, 1000, 10):
        p.apply_async(factorial, (N, d))

    # Mark pool as closed -- no more tasks can be added.
    p.close()

    # Wait for tasks to exit
    p.join()

    # Output results
    for k, v in sorted(d.items()):
        print k, v

This will produce the output:

1 1
11 39916800
21 51090942171709440000
31 8222838654177922817725562880000000
41 33452526613163807108170062053440751665152000000000
51 15511187532873822802242430164693032110632597200169861120000...

PEP 3101: Advanced String Formatting

In Python 3.0, the % operator is supplemented by a more powerful string formatting method, format(). Support for the str.format() method has been backported to Python 2.6.

In 2.6, both 8-bit and Unicode strings have a .format() method that treats the string as a template and takes the arguments to be formatted. The formatting template uses curly brackets ({, }) as special characters:

>>> # Substitute positional argument 0 into the string.
>>> "User ID: {0}".format("root")
'User ID: root'
>>> # Use the named keyword arguments
>>> "User ID: {uid}   Last seen: {last_login}".format(
...    uid="root",
...    last_login = "5 Mar 2008 07:20")
'User ID: root   Last seen: 5 Mar 2008 07:20'

Curly brackets can be escaped by doubling them:

>>> "Empty dict: {{}}".format()
"Empty dict: {}"

Field names can be integers indicating positional arguments, such as {0}, {1}, etc. or names of keyword arguments. You can also supply compound field names that read attributes or access dictionary keys:

>>> import sys
>>> print 'Platform: {0.platform}\nPython version: {0.version}'.format(sys)
Platform: darwin
Python version: 2.6a1+ (trunk:61261M, Mar  5 2008, 20:29:41)
[GCC 4.0.1 (Apple Computer, Inc. build 5367)]'

>>> import mimetypes
>>> 'Content-type: {0[.mp4]}'.format(mimetypes.types_map)
'Content-type: video/mp4'

Note that when using dictionary-style notation such as [.mp4], you don’t need to put any quotation marks around the string; it will look up the value using .mp4 as the key. Strings beginning with a number will be converted to an integer. You can’t write more complicated expressions inside a format string.

So far we’ve shown how to specify which field to substitute into the resulting string. The precise formatting used is also controllable by adding a colon followed by a format specifier. For example:

>>> # Field 0: left justify, pad to 15 characters
>>> # Field 1: right justify, pad to 6 characters
>>> fmt = '{0:15} ${1:>6}'
>>> fmt.format('Registration', 35)
'Registration    $    35'
>>> fmt.format('Tutorial', 50)
'Tutorial        $    50'
>>> fmt.format('Banquet', 125)
'Banquet         $   125'

Format specifiers can reference other fields through nesting:

>>> fmt = '{0:{1}}'
>>> width = 15
>>> fmt.format('Invoice #1234', width)
'Invoice #1234  '
>>> width = 35
>>> fmt.format('Invoice #1234', width)
'Invoice #1234                      '

The alignment of a field within the desired width can be specified:

Format specifiers can also include a presentation type, which controls how the value is formatted. For example, floating-point numbers can be formatted as a general number or in exponential notation:

>>> '{0:g}'.format(3.75)
'3.75'
>>> '{0:e}'.format(3.75)
'3.750000e+00'

A variety of presentation types are available. Consult the 2.6 documentation for a complete list; here’s a sample:

Classes and types can define a __format__() method to control how they’re formatted. It receives a single argument, the format specifier:

def __format__(self, format_spec):
    if isinstance(format_spec, unicode):
        return unicode(str(self))
    else:
        return str(self)

There’s also a format() builtin that will format a single value. It calls the type’s __format__() method with the provided specifier:

>>> format(75.6564, '.2f')
'75.66'

PEP 3105: print As a Function

The print statement becomes the print() function in Python 3.0. Making print() a function makes it possible to replace the function by doing def print(...) or importing a new function from somewhere else.

Python 2.6 has a __future__ import that removes print as language syntax, letting you use the functional form instead. For example:

>>> from __future__ import print_function
>>> print('# of entries', len(dictionary), file=sys.stderr)

The signature of the new function is:

def print(*args, sep=' ', end='\n', file=None)

The parameters are:

• args: positional arguments whose values will be printed out.

• sep: the separator, which will be printed between arguments.

• end: the ending text, which will be printed after all of the arguments have been output.

• file: the file object to which the output will be sent.

PEP 3110: Exception-Handling Changes

One error that Python programmers occasionally make is writing the following code:

try:
    ...
except TypeError, ValueError:  # Wrong!
    ...

The author is probably trying to catch both TypeError and ValueError exceptions, but this code actually does something different: it will catch TypeError and bind the resulting exception object to the local name "ValueError". The ValueError exception will not be caught at all. The correct code specifies a tuple of exceptions:

try:
    ...
except (TypeError, ValueError):
    ...

This error happens because the use of the comma here is ambiguous: does it indicate two different nodes in the parse tree, or a single node that’s a tuple?

Python 3.0 makes this unambiguous by replacing the comma with the word “as”. To catch an exception and store the exception object in the variable exc, you must write:

try:
    ...
except TypeError as exc:
    ...

Python 3.0 will only support the use of “as”, and therefore interprets the first example as catching two different exceptions. Python 2.6 supports both the comma and “as”, so existing code will continue to work. We therefore suggest using “as” when writing new Python code that will only be executed with 2.6.

PEP 3112: Byte Literals

Python 3.0 adopts Unicode as the language’s fundamental string type and denotes 8-bit literals differently, either as b'string' or using a bytes constructor. For future compatibility, Python 2.6 adds bytes as a synonym for the str type, and it also supports the b'' notation.

The 2.6 str differs from 3.0’s bytes type in various ways; most notably, the constructor is completely different. In 3.0, bytes([65, 66, 67]) is 3 elements long, containing the bytes representing ABC; in 2.6, bytes([65, 66, 67]) returns the 12-byte string representing the str() of the list.

The primary use of bytes in 2.6 will be to write tests of object type such as isinstance(x, bytes). This will help the 2to3 converter, which can’t tell whether 2.x code intends strings to contain either characters or 8-bit bytes; you can now use either bytes or str to represent your intention exactly, and the resulting code will also be correct in Python 3.0.

There’s also a __future__ import that causes all string literals to become Unicode strings. This means that \u escape sequences can be used to include Unicode characters:

from __future__ import unicode_literals

s = ('\u751f\u3080\u304e\u3000\u751f\u3054'
     '\u3081\u3000\u751f\u305f\u307e\u3054')

print len(s)               # 12 Unicode characters

At the C level, Python 3.0 will rename the existing 8-bit string type, called PyStringObject in Python 2.x, to PyBytesObject. Python 2.6 uses #define to support using the names PyBytesObject(), PyBytes_Check(), PyBytes_FromStringAndSize(), and all the other functions and macros used with strings.

Instances of the bytes type are immutable just as strings are. A new bytearray type stores a mutable sequence of bytes:

>>> bytearray([65, 66, 67])
bytearray(b'ABC')
>>> b = bytearray(u'\u21ef\u3244', 'utf-8')
>>> b
bytearray(b'\xe2\x87\xaf\xe3\x89\x84')
>>> b[0] = '\xe3'
>>> b
bytearray(b'\xe3\x87\xaf\xe3\x89\x84')
>>> unicode(str(b), 'utf-8')
u'\u31ef \u3244'

Byte arrays support most of the methods of string types, such as startswith()/endswith(), find()/rfind(), and some of the methods of lists, such as append(), pop(), and reverse().

>>> b = bytearray('ABC')
>>> b.append('d')
>>> b.append(ord('e'))
>>> b
bytearray(b'ABCde')

There’s also a corresponding C API, with PyByteArray_FromObject(), PyByteArray_FromStringAndSize(), and various other functions.

PEP 3116: New I/O Library

Python’s built-in file objects support a number of methods, but file-like objects don’t necessarily support all of them. Objects that imitate files usually support read() and write(), but they may not support readline(), for example. Python 3.0 introduces a layered I/O library in the io module that separates buffering and text-handling features from the fundamental read and write operations.

There are three levels of abstract base classes provided by the io module:

• RawIOBase defines raw I/O operations: read(), readinto(), write(), seek(), tell(), truncate(), and close(). Most of the methods of this class will often map to a single system call. There are also readable(), writable(), and seekable() methods for determining what operations a given object will allow.

  Python 3.0 has concrete implementations of this class for files and sockets, but Python 2.6 hasn’t restructured its file and socket objects in this way.

• BufferedIOBase is an abstract base class that buffers data in memory to reduce the number of system calls used, making I/O processing more efficient. It supports all of the methods of RawIOBase, and adds a raw attribute holding the underlying raw object.

  There are five concrete classes implementing this ABC. BufferedWriter and BufferedReader are for objects that support write-only or read-only usage that have a seek() method for random access. BufferedRandom objects support read and write access upon the same underlying stream, and BufferedRWPair is for objects such as TTYs that have both read and write operations acting upon unconnected streams of data. The BytesIO class supports reading, writing, and seeking over an in-memory buffer.

• TextIOBase: Provides functions for reading and writing strings (remember, strings will be Unicode in Python 3.0), and supporting universal newlines. TextIOBase defines the readline() method and supports iteration upon objects.

  There are two concrete implementations. TextIOWrapper wraps a buffered I/O object, supporting all of the methods for text I/O and adding a buffer attribute for access to the underlying object. StringIO simply buffers everything in memory without ever writing anything to disk.

  (In Python 2.6, io.StringIO is implemented in pure Python, so it’s pretty slow. You should therefore stick with the existing StringIO module or cStringIO for now. At some point Python 3.0’s io module will be rewritten into C for speed, and perhaps the C implementation will be backported to the 2.x releases.)

In Python 2.6, the underlying implementations haven’t been restructured to build on top of the io module’s classes. The module is being provided to make it easier to write code that’s forward-compatible with 3.0, and to save developers the effort of writing their own implementations of buffering and text I/O.

PEP 3118: Revised Buffer Protocol

The buffer protocol is a C-level API that lets Python types exchange pointers into their internal representations. A memory-mapped file can be viewed as a buffer of characters, for example, and this lets another module such as re treat memory-mapped files as a string of characters to be searched.

The primary users of the buffer protocol are numeric-processing packages such as NumPy, which expose the internal representation of arrays so that callers can write data directly into an array instead of going through a slower API. This PEP updates the buffer protocol in light of experience from NumPy development, adding a number of new features such as indicating the shape of an array or locking a memory region.

The most important new C API function is PyObject_GetBuffer(PyObject *obj, Py_buffer *view, int flags), which takes an object and a set of flags, and fills in the Py_buffer structure with information about the object’s memory representation. Objects can use this operation to lock memory in place while an external caller could be modifying the contents, so there’s a corresponding PyBuffer_Release(Py_buffer *view) to indicate that the external caller is done.

The flags argument to PyObject_GetBuffer() specifies constraints upon the memory returned. Some examples are:

• PyBUF_WRITABLE indicates that the memory must be writable.

• PyBUF_LOCK requests a read-only or exclusive lock on the memory.

• PyBUF_C_CONTIGUOUS and PyBUF_F_CONTIGUOUS requests a C-contiguous (last dimension varies the fastest) or Fortran-contiguous (first dimension varies the fastest) array layout.

Two new argument codes for PyArg_ParseTuple(), s* and z*, return locked buffer objects for a parameter.

PEP 3119: Abstract Base Classes

Some object-oriented languages such as Java support interfaces, declaring that a class has a given set of methods or supports a given access protocol. Abstract Base Classes (or ABCs) are an equivalent feature for Python. The ABC support consists of an abc module containing a metaclass called ABCMeta, special handling of this metaclass by the isinstance() and issubclass() builtins, and a collection of basic ABCs that the Python developers think will be widely useful. Future versions of Python will probably add more ABCs.

Let’s say you have a particular class and wish to know whether it supports dictionary-style access. The phrase “dictionary-style” is vague, however. It probably means that accessing items with obj[1] works. Does it imply that setting items with obj[2] = value works? Or that the object will have keys(), values(), and items() methods? What about the iterative variants such as iterkeys()? copy() and update()? Iterating over the object with iter()?

The Python 2.6 collections module includes a number of different ABCs that represent these distinctions. Iterable indicates that a class defines __iter__(), and Container means the class defines a __contains__() method and therefore supports x in y expressions. The basic dictionary interface of getting items, setting items, and keys(), values(), and items(), is defined by the MutableMapping ABC.

You can derive your own classes from a particular ABC to indicate they support that ABC’s interface:

import collections

class Storage(collections.MutableMapping):
    ...

Alternatively, you could write the class without deriving from the desired ABC and instead register the class by calling the ABC’s register() method:

import collections

class Storage:
    ...

collections.MutableMapping.register(Storage)

For classes that you write, deriving from the ABC is probably clearer. The register() method is useful when you’ve written a new ABC that can describe an existing type or class, or if you want to declare that some third-party class implements an ABC. For example, if you defined a PrintableType ABC, it’s legal to do:

# Register Python's types
PrintableType.register(int)
PrintableType.register(float)
PrintableType.register(str)

Classes should obey the semantics specified by an ABC, but Python can’t check this; it’s up to the class author to understand the ABC’s requirements and to implement the code accordingly.

To check whether an object supports a particular interface, you can now write:

def func(d):
    if not isinstance(d, collections.MutableMapping):
        raise ValueError("Mapping object expected, not %r" % d)

Don’t feel that you must now begin writing lots of checks as in the above example. Python has a strong tradition of duck-typing, where explicit type-checking is never done and code simply calls methods on an object, trusting that those methods will be there and raising an exception if they aren’t. Be judicious in checking for ABCs and only do it where it’s absolutely necessary.

You can write your own ABCs by using abc.ABCMeta as the metaclass in a class definition:

from abc import ABCMeta, abstractmethod

class Drawable():
    __metaclass__ = ABCMeta

    @abstractmethod
    def draw(self, x, y, scale=1.0):
        pass

    def draw_doubled(self, x, y):
        self.draw(x, y, scale=2.0)

class Square(Drawable):
    def draw(self, x, y, scale):
        ...

In the Drawable ABC above, the draw_doubled() method renders the object at twice its size and can be implemented in terms of other methods described in Drawable. Classes implementing this ABC therefore don’t need to provide their own implementation of draw_doubled(), though they can do so. An implementation of draw() is necessary, though; the ABC can’t provide a useful generic implementation.

You can apply the @abstractmethod decorator to methods such as draw() that must be implemented; Python will then raise an exception for classes that don’t define the method. Note that the exception is only raised when you actually try to create an instance of a subclass lacking the method:

>>> class Circle(Drawable):
...     pass
...
>>> c = Circle()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: Can't instantiate abstract class Circle with abstract methods draw
>>>

Abstract data attributes can be declared using the @abstractproperty decorator:

from abc import abstractproperty
...

@abstractproperty
def readonly(self):
   return self._x

Subclasses must then define a readonly() property.

PEP 3127: Integer Literal Support and Syntax

Python 3.0 changes the syntax for octal (base-8) integer literals, prefixing them with “0o” or “0O” instead of a leading zero, and adds support for binary (base-2) integer literals, signalled by a “0b” or “0B” prefix.

Python 2.6 doesn’t drop support for a leading 0 signalling an octal number, but it does add support for “0o” and “0b”:

>>> 0o21, 2*8 + 1
(17, 17)
>>> 0b101111
47

The oct() builtin still returns numbers prefixed with a leading zero, and a new bin() builtin returns the binary representation for a number:

>>> oct(42)
'052'
>>> future_builtins.oct(42)
'0o52'
>>> bin(173)
'0b10101101'

The int() and long() builtins will now accept the “0o” and “0b” prefixes when base-8 or base-2 are requested, or when the base argument is zero (signalling that the base used should be determined from the string):

>>> int ('0o52', 0)
42
>>> int('1101', 2)
13
>>> int('0b1101', 2)
13
>>> int('0b1101', 0)
13

PEP 3129: Class Decorators

Decorators have been extended from functions to classes. It’s now legal to write:

@foo
@bar
class A:
  pass

This is equivalent to:

class A:
  pass

A = foo(bar(A))

PEP 3141: A Type Hierarchy for Numbers

Python 3.0 adds several abstract base classes for numeric types inspired by Scheme’s numeric tower. These classes were backported to 2.6 as the numbers module.

The most general ABC is Number. It defines no operations at all, and only exists to allow checking if an object is a number by doing isinstance(obj, Number).

Complex is a subclass of Number. Complex numbers can undergo the basic operations of addition, subtraction, multiplication, division, and exponentiation, and you can retrieve the real and imaginary parts and obtain a number’s conjugate. Python’s built-in complex type is an implementation of Complex.

Real further derives from Complex, and adds operations that only work on real numbers: floor(), trunc(), rounding, taking the remainder mod N, floor division, and comparisons.

Rational numbers derive from Real, have numerator and denominator properties, and can be converted to floats. Python 2.6 adds a simple rational-number class, Fraction, in the fractions module. (It’s called Fraction instead of Rational to avoid a name clash with numbers.Rational.)

Integral numbers derive from Rational, and can be shifted left and right with << and >>, combined using bitwise operations such as & and |, and can be used as array indexes and slice boundaries.

In Python 3.0, the PEP slightly redefines the existing builtins round(), math.floor(), math.ceil(), and adds a new one, math.trunc(), that’s been backported to Python 2.6. math.trunc() rounds toward zero, returning the closest Integral that’s between the function’s argument and zero.

>>> from fractions import Fraction
>>> a = Fraction(2, 3)
>>> b = Fraction(2, 5)
>>> float(a), float(b)
(0.66666666666666663, 0.40000000000000002)
>>> a+b
Fraction(16, 15)
>>> a/b
Fraction(5, 3)

>>> (2.5) .as_integer_ratio()
(5, 2)
>>> (3.1415) .as_integer_ratio()
(7074029114692207L, 2251799813685248L)
>>> (1./3) .as_integer_ratio()
(6004799503160661L, 18014398509481984L)

Other Language Changes

Some smaller changes made to the core Python language are:

• Directories and zip archives containing a __main__.py file can now be executed directly by passing their name to the interpreter. The directory or zip archive is automatically inserted as the first entry in sys.path. (Suggestion and initial patch by Andy Chu, subsequently revised by Phillip J. Eby and Nick Coghlan; bpo-1739468 .)

• The hasattr() function was catching and ignoring all errors, under the assumption that they meant a __getattr__() method was failing somehow and the return value of hasattr() would therefore be False. This logic shouldn’t be applied to KeyboardInterrupt and SystemExit, however; Python 2.6 will no longer discard such exceptions when hasattr() encounters them. (Fixed by Benjamin Peterson; bpo-2196 .)

• When calling a function using the ** syntax to provide keyword arguments, you are no longer required to use a Python dictionary; any mapping will now work:

  >>> def f(**kw):
  ...    print sorted(kw)
  ...
  >>> ud=UserDict.UserDict()
  >>> ud['a'] = 1
  >>> ud['b'] = 'string'
  >>> f(**ud)
  ['a', 'b']
  

  (Contributed by Alexander Belopolsky; bpo-1686487 .)

  It’s also become legal to provide keyword arguments after a *args argument to a function call.

  >>> def f(*args, **kw):
  ...     print args, kw
  ...
  >>> f(1,2,3, *(4,5,6), keyword=13)
  (1, 2, 3, 4, 5, 6) {'keyword': 13}
  

  Previously this would have been a syntax error. (Contributed by Amaury Forgeot d’Arc; bpo-3473 .)

• A new builtin, next(iterator, [default]) returns the next item from the specified iterator. If the default argument is supplied, it will be returned if iterator has been exhausted; otherwise, the StopIteration exception will be raised. (Backported in bpo-2719 .)

• Tuples now have index() and count() methods matching the list type’s index() and count() methods:

  >>> t = (0,1,2,3,4,0,1,2)
  >>> t.index(3)
  3
  >>> t.count(0)
  2
  

  (Contributed by Raymond Hettinger)

• The built-in types now have improved support for extended slicing syntax, accepting various combinations of (start, stop, step). Previously, the support was partial and certain corner cases wouldn’t work. (Implemented by Thomas Wouters.)

• Properties now have three attributes, getter, setter and deleter, that are decorators providing useful shortcuts for adding a getter, setter or deleter function to an existing property. You would use them like this:

  class C(object):
      @property
      def x(self):
          return self._x
  
      @x.setter
      def x(self, value):
          self._x = value
  
      @x.deleter
      def x(self):
          del self._x
  
  class D(C):
      @C.x.getter
      def x(self):
          return self._x * 2
  
      @x.setter
      def x(self, value):
          self._x = value / 2
  

• Several methods of the built-in set types now accept multiple iterables: intersection(), intersection_update(), union(), update(), difference() and difference_update().

  >>> s=set('1234567890')
  >>> s.intersection('abc123', 'cdf246')  # Intersection between all inputs
  set(['2'])
  >>> s.difference('246', '789')
  set(['1', '0', '3', '5'])
  

  (Contributed by Raymond Hettinger.)

• Many floating-point features were added. The float() function will now turn the string nan into an IEEE 754 Not A Number value, and +inf and -inf into positive or negative infinity. This works on any platform with IEEE 754 semantics. (Contributed by Christian Heimes; bpo-1635 .)

  Other functions in the math module, isinf() and isnan(), return true if their floating-point argument is infinite or Not A Number. (bpo-1640 )

  Conversion functions were added to convert floating-point numbers into hexadecimal strings (bpo-3008 ). These functions convert floats to and from a string representation without introducing rounding errors from the conversion between decimal and binary. Floats have a hex() method that returns a string representation, and the float.fromhex() method converts a string back into a number:

  >>> a = 3.75
  >>> a.hex()
  '0x1.e000000000000p+1'
  >>> float.fromhex('0x1.e000000000000p+1')
  3.75
  >>> b=1./3
  >>> b.hex()
  '0x1.5555555555555p-2'
  

• A numerical nicety: when creating a complex number from two floats on systems that support signed zeros (-0 and +0), the complex() constructor will now preserve the sign of the zero. (Fixed by Mark T. Dickinson; bpo-1507 .)

• Classes that inherit a __hash__() method from a parent class can set __hash__ = None to indicate that the class isn’t hashable. This will make hash(obj) raise a TypeError and the class will not be indicated as implementing the Hashable ABC.

  You should do this when you’ve defined a __cmp__() or __eq__() method that compares objects by their value rather than by identity. All objects have a default hash method that uses id(obj) as the hash value. There’s no tidy way to remove the __hash__() method inherited from a parent class, so assigning None was implemented as an override. At the C level, extensions can set tp_hash to PyObject_HashNotImplemented(). (Fixed by Nick Coghlan and Amaury Forgeot d’Arc; bpo-2235 .)

• The GeneratorExit exception now subclasses BaseException instead of Exception. This means that an exception handler that does except Exception: will not inadvertently catch GeneratorExit. (Contributed by Chad Austin; bpo-1537 .)

• Generator objects now have a gi_code attribute that refers to the original code object backing the generator. (Contributed by Collin Winter; bpo-1473257 .)

• The compile() built-in function now accepts keyword arguments as well as positional parameters. (Contributed by Thomas Wouters; bpo-1444529 .)

• The complex() constructor now accepts strings containing parenthesized complex numbers, meaning that complex(repr(cplx)) will now round-trip values. For example, complex('(3+4j)') now returns the value (3+4j). (bpo-1491866 )

• The string translate() method now accepts None as the translation table parameter, which is treated as the identity transformation. This makes it easier to carry out operations that only delete characters. (Contributed by Bengt Richter and implemented by Raymond Hettinger; bpo-1193128 .)

• The built-in dir() function now checks for a __dir__() method on the objects it receives. This method must return a list of strings containing the names of valid attributes for the object, and lets the object control the value that dir() produces. Objects that have __getattr__() or __getattribute__() methods can use this to advertise pseudo-attributes they will honor. (bpo-1591665 )

• Instance method objects have new attributes for the object and function comprising the method; the new synonym for im_self is __self__, and im_func is also available as __func__. The old names are still supported in Python 2.6, but are gone in 3.0.

• An obscure change: when you use the locals() function inside a class statement, the resulting dictionary no longer returns free variables. (Free variables, in this case, are variables referenced in the class statement that aren’t attributes of the class.)

New and Improved Modules

As in every release, Python’s standard library received a number of enhancements and bug fixes. Here’s a partial list of the most notable changes, sorted alphabetically by module name. Consult the Misc/NEWS file in the source tree for a more complete list of changes, or look through the Subversion logs for all the details.

• The asyncore and asynchat modules are being actively maintained again, and a number of patches and bugfixes were applied. (Maintained by Josiah Carlson; see bpo-1736190 for one patch.)

• The bsddb module also has a new maintainer, Jesús Cea Avion, and the package is now available as a standalone package. The web page for the package is www.jcea.es/programacion/pybsddb.htm . The plan is to remove the package from the standard library in Python 3.0, because its pace of releases is much more frequent than Python’s.

  The bsddb.dbshelve module now uses the highest pickling protocol available, instead of restricting itself to protocol 1. (Contributed by W. Barnes.)

• The cgi module will now read variables from the query string of an HTTP POST request. This makes it possible to use form actions with URLs that include query strings such as “/cgi-bin/add.py?category=1”. (Contributed by Alexandre Fiori and Nubis; bpo-1817 .)

  The parse_qs() and parse_qsl() functions have been relocated from the cgi module to the urlparse module. The versions still available in the cgi module will trigger PendingDeprecationWarning messages in 2.6 (bpo-600362 ).

• The cmath module underwent extensive revision, contributed by Mark Dickinson and Christian Heimes. Five new functions were added:

  • polar() converts a complex number to polar form, returning the modulus and argument of the complex number.

  • rect() does the opposite, turning a modulus, argument pair back into the corresponding complex number.

  • phase() returns the argument (also called the angle) of a complex number.

  • isnan() returns True if either the real or imaginary part of its argument is a NaN.

  • isinf() returns True if either the real or imaginary part of its argument is infinite.

  The revisions also improved the numerical soundness of the cmath module. For all functions, the real and imaginary parts of the results are accurate to within a few units of least precision (ulps) whenever possible. See bpo-1381 for the details. The branch cuts for asinh(), atanh(): and atan() have also been corrected.

  The tests for the module have been greatly expanded; nearly 2000 new test cases exercise the algebraic functions.

  On IEEE 754 platforms, the cmath module now handles IEEE 754 special values and floating-point exceptions in a manner consistent with Annex ‘G’ of the C99 standard.

• A new data type in the collections module: namedtuple(typename, fieldnames) is a factory function that creates subclasses of the standard tuple whose fields are accessible by name as well as index. For example:

  >>> var_type = collections.namedtuple('variable',
  ...             'id name type size')
  >>> # Names are separated by spaces or commas.
  >>> # 'id, name, type, size' would also work.
  >>> var_type._fields
  ('id', 'name', 'type', 'size')
  
  >>> var = var_type(1, 'frequency', 'int', 4)
  >>> print var[0], var.id    # Equivalent
  1 1
  >>> print var[2], var.type  # Equivalent
  int int
  >>> var._asdict()
  {'size': 4, 'type': 'int', 'id': 1, 'name': 'frequency'}
  >>> v2 = var._replace(name='amplitude')
  >>> v2
  variable(id=1, name='amplitude', type='int', size=4)
  

  Several places in the standard library that returned tuples have been modified to return namedtuple instances. For example, the Decimal.as_tuple() method now returns a named tuple with sign, digits, and exponent fields.

  (Contributed by Raymond Hettinger.)

• Another change to the collections module is that the deque type now supports an optional maxlen parameter; if supplied, the deque’s size will be restricted to no more than maxlen items. Adding more items to a full deque causes old items to be discarded.

  >>> from collections import deque
  >>> dq=deque(maxlen=3)
  >>> dq
  deque([], maxlen=3)
  >>> dq.append(1); dq.append(2); dq.append(3)
  >>> dq
  deque([1, 2, 3], maxlen=3)
  >>> dq.append(4)
  >>> dq
  deque([2, 3, 4], maxlen=3)
  

  (Contributed by Raymond Hettinger.)

• The Cookie module’s Morsel objects now support an httponly attribute. In some browsers. cookies with this attribute set cannot be accessed or manipulated by JavaScript code. (Contributed by Arvin Schnell; bpo-1638033 .)

• A new window method in the curses module, chgat(), changes the display attributes for a certain number of characters on a single line. (Contributed by Fabian Kreutz.)

  # Boldface text starting at y=0,x=21
  # and affecting the rest of the line.
  stdscr.chgat(0, 21, curses.A_BOLD)
  

  The Textbox class in the curses.textpad module now supports editing in insert mode as well as overwrite mode. Insert mode is enabled by supplying a true value for the insert_mode parameter when creating the Textbox instance.

• The datetime module’s strftime() methods now support a %f format code that expands to the number of microseconds in the object, zero-padded on the left to six places. (Contributed by Skip Montanaro; bpo-1158 .)

• The decimal module was updated to version 1.66 of the General Decimal Specification . New features include some methods for some basic mathematical functions such as exp() and log10():

  >>> Decimal(1).exp()
  Decimal("2.718281828459045235360287471")
  >>> Decimal("2.7182818").ln()
  Decimal("0.9999999895305022877376682436")
  >>> Decimal(1000).log10()
  Decimal("3")
  

  The as_tuple() method of Decimal objects now returns a named tuple with sign, digits, and exponent fields.

  (Implemented by Facundo Batista and Mark Dickinson. Named tuple support added by Raymond Hettinger.)

• The difflib module’s SequenceMatcher class now returns named tuples representing matches, with a, b, and size attributes. (Contributed by Raymond Hettinger.)

• An optional timeout parameter, specifying a timeout measured in seconds, was added to the ftplib.FTP class constructor as well as the connect() method. (Added by Facundo Batista.) Also, the FTP class’s storbinary() and storlines() now take an optional callback parameter that will be called with each block of data after the data has been sent. (Contributed by Phil Schwartz; bpo-1221598 .)

• The reduce() built-in function is also available in the functools module. In Python 3.0, the builtin has been dropped and reduce() is only available from functools; currently there are no plans to drop the builtin in the 2.x series. (Patched by Christian Heimes; bpo-1739906 .)

• When possible, the getpass module will now use /dev/tty to print a prompt message and read the password, falling back to standard error and standard input. If the password may be echoed to the terminal, a warning is printed before the prompt is displayed. (Contributed by Gregory P. Smith.)

• The glob.glob() function can now return Unicode filenames if a Unicode path was used and Unicode filenames are matched within the directory. (bpo-1001604 )

• A new function in the heapq module, merge(iter1, iter2, ...), takes any number of iterables returning data in sorted order, and returns a new generator that returns the contents of all the iterators, also in sorted order. For example:

  >>> list(heapq.merge([1, 3, 5, 9], [2, 8, 16]))
  [1, 2, 3, 5, 8, 9, 16]
  

  Another new function, heappushpop(heap, item), pushes item onto heap, then pops off and returns the smallest item. This is more efficient than making a call to heappush() and then heappop().

  heapq is now implemented to only use less-than comparison, instead of the less-than-or-equal comparison it previously used. This makes