What’s New In Python 3.2


Raymond Hettinger

This article explains the new features in Python 3.2 as compared to 3.1. It focuses on a few highlights and gives a few examples. For full details, see the Misc/NEWS file.

See also

PEP 392 - Python 3.2 Release Schedule

PEP 384: Defining a Stable ABI

In the past, extension modules built for one Python version were often not usable with other Python versions. Particularly on Windows, every feature release of Python required rebuilding all extension modules that one wanted to use. This requirement was the result of the free access to Python interpreter internals that extension modules could use.

With Python 3.2, an alternative approach becomes available: extension modules which restrict themselves to a limited API (by defining Py_LIMITED_API) cannot use many of the internals, but are constrained to a set of API functions that are promised to be stable for several releases. As a consequence, extension modules built for 3.2 in that mode will also work with 3.3, 3.4, and so on. Extension modules that make use of details of memory structures can still be built, but will need to be recompiled for every feature release.

See also

PEP 384 - Defining a Stable ABI

PEP written by Martin von Löwis.

PEP 389: Argparse Command Line Parsing Module

A new module for command line parsing, argparse, was introduced to overcome the limitations of optparse which did not provide support for positional arguments (not just options), subcommands, required options and other common patterns of specifying and validating options.

This module has already had widespread success in the community as a third-party module. Being more fully featured than its predecessor, the argparse module is now the preferred module for command-line processing. The older module is still being kept available because of the substantial amount of legacy code that depends on it.

Here’s an annotated example parser showing features like limiting results to a set of choices, specifying a metavar in the help screen, validating that one or more positional arguments is present, and making a required option:

import argparse
parser = argparse.ArgumentParser(
            description = 'Manage servers',         # main description for help
            epilog = 'Tested on Solaris and Linux') # displayed after help
parser.add_argument('action',                       # argument name
            choices = ['deploy', 'start', 'stop'],  # three allowed values
            help = 'action on each target')         # help msg
            metavar = 'HOSTNAME',                   # var name used in help msg
            nargs = '+',                            # require one or more targets
            help = 'url for target machines')       # help msg explanation
parser.add_argument('-u', '--user',                 # -u or --user option
            required = True,                        # make it a required argument
            help = 'login as user')

Example of calling the parser on a command string:

>>> cmd = 'deploy sneezy.example.com sleepy.example.com -u skycaptain'
>>> result = parser.parse_args(cmd.split())
>>> result.action
>>> result.targets
['sneezy.example.com', 'sleepy.example.com']
>>> result.user

Example of the parser’s automatically generated help:

>>> parser.parse_args('-h'.split())

usage: manage_cloud.py [-h] -u USER
                       {deploy,start,stop} HOSTNAME [HOSTNAME ...]

Manage servers

positional arguments:
  {deploy,start,stop}   action on each target
  HOSTNAME              url for target machines

optional arguments:
  -h, --help            show this help message and exit
  -u USER, --user USER  login as user

Tested on Solaris and Linux

An especially nice argparse feature is the ability to define subparsers, each with their own argument patterns and help displays:

import argparse
parser = argparse.ArgumentParser(prog='HELM')
subparsers = parser.add_subparsers()

parser_l = subparsers.add_parser('launch', help='Launch Control')   # first subgroup
parser_l.add_argument('-m', '--missiles', action='store_true')
parser_l.add_argument('-t', '--torpedos', action='store_true')

parser_m = subparsers.add_parser('move', help='Move Vessel',        # second subgroup
                                 aliases=('steer', 'turn'))         # equivalent names
parser_m.add_argument('-c', '--course', type=int, required=True)
parser_m.add_argument('-s', '--speed', type=int, default=0)
$ ./helm.py --help                         # top level help (launch and move)
$ ./helm.py launch --help                  # help for launch options
$ ./helm.py launch --missiles              # set missiles=True and torpedos=False
$ ./helm.py steer --course 180 --speed 5   # set movement parameters

See also

PEP 389 - New Command Line Parsing Module

PEP written by Steven Bethard.

Upgrading optparse code for details on the differences from optparse.

PEP 391: Dictionary Based Configuration for Logging

The logging module provided two kinds of configuration, one style with function calls for each option or another style driven by an external file saved in a ConfigParser format. Those options did not provide the flexibility to create configurations from JSON or YAML files, nor did they support incremental configuration, which is needed for specifying logger options from a command line.

To support a more flexible style, the module now offers logging.config.dictConfig() for specifying logging configuration with plain Python dictionaries. The configuration options include formatters, handlers, filters, and loggers. Here’s a working example of a configuration dictionary:

{"version": 1,
 "formatters": {"brief": {"format": "%(levelname)-8s: %(name)-15s: %(message)s"},
                "full": {"format": "%(asctime)s %(name)-15s %(levelname)-8s %(message)s"}
 "handlers": {"console": {
                   "class": "logging.StreamHandler",
                   "formatter": "brief",
                   "level": "INFO",
                   "stream": "ext://sys.stdout"},
              "console_priority": {
                   "class": "logging.StreamHandler",
                   "formatter": "full",
                   "level": "ERROR",
                   "stream": "ext://sys.stderr"}
 "root": {"level": "DEBUG", "handlers": ["console", "console_priority"]}}

If that dictionary is stored in a file called conf.json, it can be loaded and called with code like this:

>>> import json, logging.config
>>> with open('conf.json') as f:
...     conf = json.load(f)
>>> logging.config.dictConfig(conf)
>>> logging.info("Transaction completed normally")
INFO    : root           : Transaction completed normally
>>> logging.critical("Abnormal termination")
2011-02-17 11:14:36,694 root            CRITICAL Abnormal termination

See also

PEP 391 - Dictionary Based Configuration for Logging

PEP written by Vinay Sajip.

PEP 3148: The concurrent.futures module

Code for creating and managing concurrency is being collected in a new top-level namespace, concurrent. Its first member is a futures package which provides a uniform high-level interface for managing threads and processes.

The design for concurrent.futures was inspired by the java.util.concurrent package. In that model, a running call and its result are represented by a Future object that abstracts features common to threads, processes, and remote procedure calls. That object supports status checks (running or done), timeouts, cancellations, adding callbacks, and access to results or exceptions.

The primary offering of the new module is a pair of executor classes for launching and managing calls. The goal of the executors is to make it easier to use existing tools for making parallel calls. They save the effort needed to setup a pool of resources, launch the calls, create a results queue, add time-out handling, and limit the total number of threads, processes, or remote procedure calls.

Ideally, each application should share a single executor across multiple components so that process and thread limits can be centrally managed. This solves the design challenge that arises when each component has its own competing strategy for resource management.

Both classes share a common interface with three methods: submit() for scheduling a callable and returning a Future object; map() for scheduling many asynchronous calls at a time, and shutdown() for freeing resources. The class is a context manager and can be used in a with statement to assure that resources are automatically released when currently pending futures are done executing.

A simple of example of ThreadPoolExecutor is a launch of four parallel threads for copying files:

import concurrent.futures, shutil
with concurrent.futures.ThreadPoolExecutor(max_workers=4) as e:
    e.submit(shutil.copy, 'src1.txt', 'dest1.txt')
    e.submit(shutil.copy, 'src2.txt', 'dest2.txt')
    e.submit(shutil.copy, 'src3.txt', 'dest3.txt')
    e.submit(shutil.copy, 'src3.txt', 'dest4.txt')

See also

PEP 3148 - Futures – Execute Computations Asynchronously

PEP written by Brian Quinlan.

Code for Threaded Parallel URL reads, an example using threads to fetch multiple web pages in parallel.

Code for computing prime numbers in parallel, an example demonstrating ProcessPoolExecutor.

PEP 3147: PYC Repository Directories

Python’s scheme for caching bytecode in .pyc files did not work well in environments with multiple Python interpreters. If one interpreter encountered a cached file created by another interpreter, it would recompile the source and overwrite the cached file, thus losing the benefits of caching.

The issue of “pyc fights” has become more pronounced as it has become commonplace for Linux distributions to ship with multiple versions of Python. These conflicts also arise with CPython alternatives such as Unladen Swallow.

To solve this problem, Python’s import machinery has been extended to use distinct filenames for each interpreter. Instead of Python 3.2 and Python 3.3 and Unladen Swallow each competing for a file called “mymodule.pyc”, they will now look for “mymodule.cpython-32.pyc”, “mymodule.cpython-33.pyc”, and “mymodule.unladen10.pyc”. And to prevent all of these new files from cluttering source directories, the pyc files are now collected in a “__pycache__” directory stored under the package directory.

Aside from the filenames and target directories, the new scheme has a few aspects that are visible to the programmer:

  • Imported modules now have a __cached__ attribute which stores the name of the actual file that was imported:

    >>> import collections
    >>> collections.__cached__ 
  • The tag that is unique to each interpreter is accessible from the imp module:

    >>> import imp
    >>> imp.get_tag() 
  • Scripts that try to deduce source filename from the imported file now need to be smarter. It is no longer sufficient to simply strip the “c” from a “.pyc” filename. Instead, use the new functions in the imp module:

    >>> imp.source_from_cache('c:/py32/lib/__pycache__/collections.cpython-32.pyc')
    >>> imp.cache_from_source('c:/py32/lib/collections.py') 
  • The py_compile and compileall modules have been updated to reflect the new naming convention and target directory. The command-line invocation of compileall has new options: -i for specifying a list of files and directories to compile and -b which causes bytecode files to be written to their legacy location rather than __pycache__.

  • The importlib.abc module has been updated with new abstract base classes for loading bytecode files. The obsolete ABCs, PyLoader and PyPycLoader, have been deprecated (instructions on how to stay Python 3.1 compatible are included with the documentation).

See also

PEP 3147 - PYC Repository Directories

PEP written by Barry Warsaw.

PEP 3149: ABI Version Tagged .so Files

The PYC repository directory allows multiple bytecode cache files to be co-located. This PEP implements a similar mechanism for shared object files by giving them a common directory and distinct names for each version.

The common directory is “pyshared” and the file names are made distinct by identifying the Python implementation (such as CPython, PyPy, Jython, etc.), the major and minor version numbers, and optional build flags (such as “d” for debug, “m” for pymalloc, “u” for wide-unicode). For an arbitrary package “foo”, you may see these files when the distribution package is installed:


In Python itself, the tags are accessible from functions in the sysconfig module:

>>> import sysconfig
>>> sysconfig.get_config_var('SOABI')       # find the version tag
>>> sysconfig.get_config_var('EXT_SUFFIX')  # find the full filename extension

See also

PEP 3149 - ABI Version Tagged .so Files

PEP written by Barry Warsaw.

PEP 3333: Python Web Server Gateway Interface v1.0.1

This informational PEP clarifies how bytes/text issues are to be handled by the WSGI protocol. The challenge is that string handling in Python 3 is most conveniently handled with the str type even though the HTTP protocol is itself bytes oriented.

The PEP differentiates so-called native strings that are used for request/response headers and metadata versus byte strings which are used for the bodies of requests and responses.

The native strings are always of type str but are restricted to code points between U+0000 through U+00FF which are translatable to bytes using Latin-1 encoding. These strings are used for the keys and values in the environment dictionary and for response headers and statuses in the start_response() function. They must follow RFC 2616 with respect to encoding. That is, they must either be ISO-8859-1 characters or use RFC 2047 MIME encoding.

For developers porting WSGI applications from Python 2, here are the salient points:

  • If the app already used strings for headers in Python 2, no change is needed.

  • If instead, the app encoded output headers or decoded input headers, then the headers will need to be re-encoded to Latin-1. For example, an output header encoded in utf-8 was using h.encode('utf-8') now needs to convert from bytes to native strings using h.encode('utf-8').decode('latin-1').

  • Values yielded by an application or sent using the write() method must be byte strings. The start_response() function and environ must use native strings. The two cannot be mixed.

For server implementers writing CGI-to-WSGI pathways or other CGI-style protocols, the users must to be able access the environment using native strings even though the underlying platform may have a different convention. To bridge this gap, the wsgiref module has a new function, wsgiref.handlers.read_environ() for transcoding CGI variables from os.environ into native strings and returning a new dictionary.

See also

PEP 3333 - Python Web Server Gateway Interface v1.0.1

PEP written by Phillip Eby.

Other Language Changes

Some smaller changes made to the core Python language are:

  • String formatting for format() and str.format() gained new capabilities for the format character #. Previously, for integers in binary, octal, or hexadecimal, it caused the output to be prefixed with ‘0b’, ‘0o’, or ‘0x’ respectively. Now it can also handle floats, complex, and Decimal, causing the output to always have a decimal point even when no digits follow it.

    >>> format(20, '#o')
    >>> format(12.34, '#5.0f')
    '  12.'

    (Suggested by Mark Dickinson and implemented by Eric Smith in bpo-7094 .)

  • There is also a new str.format_map() method that extends the capabilities of the existing str.format() method by accepting arbitrary mapping objects. This new method makes it possible to use string formatting with any of Python’s many dictionary-like objects such as defaultdict, Shelf, ConfigParser, or dbm. It is also useful with custom dict subclasses that normalize keys before look-up or that supply a __missing__() method for unknown keys:

    >>> import shelve
    >>> d = shelve.open('tmp.shl')
    >>> 'The {project_name} status is {status} as of {date}'.format_map(d)
    'The testing project status is green as of February 15, 2011'
    >>> class LowerCasedDict(dict):
    ...     def __getitem__(self, key):
    ...         return dict.__getitem__(self, key.lower())
    >>> lcd = LowerCasedDict(part='widgets', quantity=10)
    >>> 'There are {QUANTITY} {Part} in stock'.format_map(lcd)
    'There are 10 widgets in stock'
    >>> class PlaceholderDict(dict):
    ...     def __missing__(self, key):
    ...         return '<{}>'.format(key)
    >>> 'Hello {name}, welcome to {location}'.format_map(PlaceholderDict())
    'Hello <name>, welcome to <location>'

(Suggested by Raymond Hettinger and implemented by Eric Smith in bpo-6081 .)

  • The interpreter can now be started with a quiet option, -q, to prevent the copyright and version information from being displayed in the interactive mode. The option can be introspected using the sys.flags attribute:

    $ python -q
    >>> sys.flags
    sys.flags(debug=0, division_warning=0, inspect=0, interactive=0,
    optimize=0, dont_write_bytecode=0, no_user_site=0, no_site=0,
    ignore_environment=0, verbose=0, bytes_warning=0, quiet=1)

    (Contributed by Marcin Wojdyr in bpo-1772833 ).

  • The hasattr() function works by calling getattr() and detecting whether an exception is raised. This technique allows it to detect methods created dynamically by __getattr__() or __getattribute__() which would otherwise be absent from the class dictionary. Formerly, hasattr would catch any exception, possibly masking genuine errors. Now, hasattr has been tightened to only catch AttributeError and let other exceptions pass through:

    >>> class A:
    ...     @property
    ...     def f(self):
    ...         return 1 // 0
    >>> a = A()
    >>> hasattr(a, 'f')
    Traceback (most recent call last):
    ZeroDivisionError: integer division or modulo by zero

    (Discovered by Yury Selivanov and fixed by Benjamin Peterson; bpo-9666 .)

  • The str() of a float or complex number is now the same as its repr(). Previously, the str() form was shorter but that just caused confusion and is no longer needed now that the shortest possible repr() is displayed by default:

    >>> import math
    >>> repr(math.pi)
    >>> str(math.pi)

    (Proposed and implemented by Mark Dickinson; bpo-9337 .)

  • memoryview objects now have a release() method and they also now support the context management protocol. This allows timely release of any resources that were acquired when requesting a buffer from the original object.

    >>> with memoryview(b'abcdefgh') as v:
    ...     print(v.tolist())
    [97, 98, 99, 100, 101, 102, 103, 104]

    (Added by Antoine Pitrou; bpo-9757 .)

  • Previously it was illegal to delete a name from the local namespace if it occurs as a free variable in a nested block:

    def outer(x):
        def inner():
            return x
        del x

    This is now allowed. Remember that the target of an except clause is cleared, so this code which used to work with Python 2.6, raised a SyntaxError with Python 3.1 and now works again:

    def f():
        def print_error():
        except Exception as e:
            # implicit "del e" here

    (See bpo-4617 .)

  • The internal structsequence tool now creates subclasses of tuple. This means that C structures like those returned by os.stat(), time.gmtime(), and sys.version_info now work like a named tuple and now work with functions and methods that expect a tuple as an argument. This is a big step forward in making the C structures as flexible as their pure Python counterparts:

    >>> import sys
    >>> isinstance(sys.version_info, tuple)
    >>> 'Version %d.%d.%d %s(%d)' % sys.version_info 
    'Version 3.2.0 final(0)'

    (Suggested by Arfrever Frehtes Taifersar Arahesis and implemented by Benjamin Peterson in bpo-8413 .)

  • Warnings are now easier to control using the PYTHONWARNINGS environment variable as an alternative to using -W at the command line:

    $ export PYTHONWARNINGS='ignore::RuntimeWarning::,once::UnicodeWarning::'

    (Suggested by Barry Warsaw and implemented by Philip Jenvey in bpo-7301 .)

  • A new warning category, ResourceWarning, has been added. It is emitted when potential issues with resource consumption or cleanup are detected. It is silenced by default in normal release builds but can be enabled through the means provided by the warnings module, or on the command line.

    A ResourceWarning is issued at interpreter shutdown if the gc.garbage list isn’t empty, and if gc.DEBUG_UNCOLLECTABLE is set, all uncollectable objects are printed. This is meant to make the programmer aware that their code contains object finalization issues.

    A ResourceWarning is also issued when a file object is destroyed without having been explicitly closed. While the deallocator for such object ensures it closes the underlying operating system resource (usually, a file descriptor), the delay in deallocating the object could produce various issues, especially under Windows. Here is an example of enabling the warning from the command line:

    $ python -q -Wdefault
    >>> f = open("foo", "wb")
    >>> del f
    __main__:1: ResourceWarning: unclosed file <_io.BufferedWriter name='foo'>

    (Added by Antoine Pitrou and Georg Brandl in bpo-10093 and bpo-477863 .)

  • range objects now support index and count methods. This is part of an effort to make more objects fully implement the collections.Sequence abstract base class. As a result, the language will have a more uniform API. In addition, range objects now support slicing and negative indices, even with values larger than sys.maxsize. This makes range more interoperable with lists:

    >>> range(0, 100, 2).count(10)
    >>> range(0, 100, 2).index(10)
    >>> range(0, 100, 2)[5]
    >>> range(0, 100, 2)[0:5]
    range(0, 10, 2)

    (Contributed by Daniel Stutzbach in bpo-9213 , by Alexander Belopolsky in bpo-2690 , and by Nick Coghlan in bpo-10889 .)

  • The callable() builtin function from Py2.x was resurrected. It provides a concise, readable alternative to using an abstract base class in an expression like isinstance(x, collections.Callable):

    >>> callable(max)
    >>> callable(20)

    (See bpo-10518 .)

  • Python’s import mechanism can now load modules installed in directories with non-ASCII characters in the path name. This solved an aggravating problem with home directories for users with non-ASCII characters in their usernames.

(Required extensive work by Victor Stinner in bpo-9425 .)

New, Improved, and Deprecated Modules

Python’s standard library has undergone significant maintenance efforts and quality improvements.

The biggest news for Python 3.2 is that the email package, mailbox module, and nntplib modules now work correctly with the bytes/text model in Python 3. For the first time, there is correct handling of messages with mixed encodings.

Throughout the standard library, there has been more careful attention to encodings and text versus bytes issues. In particular, interactions with the operating system are now better able to exchange non-ASCII data using the Windows MBCS encoding, locale-aware encodings, or UTF-8.

Another significant win is the addition of substantially better support for SSL connections and security certificates.

In addition, more classes now implement a context manager to support convenient and reliable resource clean-up using a with statement.


The usability of the email package in Python 3 has been mostly fixed by the extensive efforts of R. David Murray. The problem was that emails are typically read and stored in the form of bytes rather than str text, and they may contain multiple encodings within a single email. So, the email package had to be extended to parse and generate email messages in bytes format.

  • New functions message_from_bytes() and message_from_binary_file(), and new classes BytesFeedParser and BytesParser allow binary message data to be parsed into model objects.

  • Given bytes input to the model, get_payload() will by default decode a message body that has a Content-Transfer-Encoding of 8bit using the charset specified in the MIME headers and return the resulting string.

  • Given bytes input to the model, Generator will convert message bodies that have a Content-Transfer-Encoding of 8bit to instead have a 7bit Content-Transfer-Encoding.

    Headers with unencoded non-ASCII bytes are deemed to be RFC 2047 -encoded using the unknown-8bit character set.

  • A new class BytesGenerator produces bytes as output, preserving any unchanged non-ASCII data that was present in the input used to build the model, including message bodies with a Content-Transfer-Encoding of 8bit.

  • The smtplib SMTP class now accepts a byte string for the msg argument to the sendmail() method, and a new method, send_message() accepts a Message object and can optionally obtain the from_addr and to_addrs addresses directly from the object.

(Proposed and implemented by R. David Murray, bpo-4661 and bpo-10321 .)


The xml.etree.ElementTree package and its xml.etree.cElementTree counterpart have been updated to version 1.3.

Several new and useful functions and methods have been added:

Two methods have been deprecated:

  • xml.etree.ElementTree.getchildren() use list(elem) instead.

  • xml.etree.ElementTree.getiterator() use Element.iter instead.

For details of the update, see Introducing ElementTree on Fredrik Lundh’s website.

(Contributed by Florent Xicluna and Fredrik Lundh, bpo-6472 .)


  • The functools module includes a new decorator for caching function calls. functools.lru_cache() can save repeated queries to an external resource whenever the results are expected to be the same.

    For example, adding a caching decorator to a database query function can save database accesses for popular searches:

    >>> import functools
    >>> @functools.lru_cache(maxsize=300)
    ... def get_phone_number(name):
    ...     c = conn.cursor()
    ...     c.execute('SELECT phonenumber FROM phonelist WHERE name=?', (name,))
    ...     return c.fetchone()[0]
    >>> for name in user_requests:        
    ...     get_phone_number(name)        # cached lookup

    To help with choosing an effective cache size, the wrapped function is instrumented for tracking cache statistics:

    >>> get_phone_number.cache_info()     
    CacheInfo(hits=4805, misses=980, maxsize=300, currsize=300)

    If the phonelist table gets updated, the outdated contents of the cache can be cleared with:

    >>> get_phone_number.cache_clear()

    (Contributed by Raymond Hettinger and incorporating design ideas from Jim Baker, Miki Tebeka, and Nick Coghlan; see recipe 498245 , recipe 577479 , bpo-10586 , and bpo-10593 .)

  • The functools.wraps() decorator now adds a __wrapped__ attribute pointing to the original callable function. This allows wrapped functions to be introspected. It also copies __annotations__ if defined. And now it also gracefully skips over missing attributes such as __doc__ which might not be defined for the wrapped callable.

    In the above example, the cache can be removed by recovering the original function:

    >>> get_phone_number = get_phone_number.__wrapped__    # uncached function

    (By Nick Coghlan and Terrence Cole; bpo-9567 , bpo-3445 , and bpo-8814 .)

  • To help write classes with rich comparison methods, a new decorator functools.total_ordering() will use existing equality and inequality methods to fill in the remaining methods.

    For example, supplying __eq__ and __lt__ will enable total_ordering() to fill-in __le__, __gt__ and __ge__:

    class Student:
        def __eq__(self, other):
            return ((self.lastname.lower(), self.firstname.lower()) ==
                    (other.lastname.lower(), other.firstname.lower()))
        def __lt__(self, other):
            return ((self.lastname.lower(), self.firstname.lower()) <
                    (other.lastname.lower(), other.firstname.lower()))

    With the total_ordering decorator, the remaining comparison methods are filled in automatically.

    (Contributed by Raymond Hettinger.)

  • To aid in porting programs from Python 2, the functools.cmp_to_key() function converts an old-style comparison function to modern key function:

    >>> # locale-aware sort order
    >>> sorted(iterable, key=cmp_to_key(locale.strcoll)) 

    For sorting examples and a brief sorting tutorial, see the Sorting HowTo tutorial.

    (Contributed by Raymond Hettinger.)


  • The itertools module has a new accumulate() function modeled on APL’s scan operator and Numpy’s accumulate function:

    >>> from itertools import accumulate
    >>> list(accumulate([8, 2, 50]))
    [8, 10, 60]
    >>> prob_dist = [0.1, 0.4, 0.2, 0.3]
    >>> list(accumulate(prob_dist))      # cumulative probability distribution
    [0.1, 0.5, 0.7, 1.0]

    For an example using accumulate(), see the examples for the random module.

    (Contributed by Raymond Hettinger and incorporating design suggestions from Mark Dickinson.)


  • The collections.Counter class now has two forms of in-place subtraction, the existing -= operator for saturating subtraction and the new subtract() method for regular subtraction. The former is suitable for multisets which only have positive counts, and the latter is more suitable for use cases that allow negative counts:

    >>> from collections import Counter
    >>> tally = Counter(dogs=5, cats=3)
    >>> tally -= Counter(dogs=2, cats=8)    # saturating subtraction
    >>> tally
    Counter({'dogs': 3})
    >>> tally = Counter(dogs=5, cats=3)
    >>> tally.subtract(dogs=2, cats=8)      # regular subtraction
    >>> tally
    Counter({'dogs': 3, 'cats': -5})

    (Contributed by Raymond Hettinger.)

  • The collections.OrderedDict class has a new method move_to_end() which takes an existing key and moves it to either the first or last position in the ordered sequence.

    The default is to move an item to the last position. This is equivalent of renewing an entry with od[k] = od.pop(k).

    A fast move-to-end operation is useful for resequencing entries. For example, an ordered dictionary can be used to track order of access by aging entries from the oldest to the most recently accessed.

    >>> from collections import OrderedDict
    >>> d = OrderedDict.fromkeys(['a', 'b', 'X', 'd', 'e'])
    >>> list(d)
    ['a', 'b', 'X', 'd', 'e']
    >>> d.move_to_end('X')
    >>> list(d)
    ['a', 'b', 'd', 'e', 'X']

    (Contributed by Raymond Hettinger.)

  • The collections.deque class grew two new methods count() and reverse() that make them more substitutable for list objects:

    >>> from collections import deque
    >>> d = deque('simsalabim')
    >>> d.count('s')
    >>> d.reverse()
    >>> d
    deque(['m', 'i', 'b', 'a', 'l', 'a', 's', 'm', 'i', 's'])

    (Contributed by Raymond Hettinger.)


The threading module has a new Barrier synchronization class for making multiple threads wait until all of them have reached a common barrier point. Barriers are useful for making sure that a task with multiple preconditions does not run until all of the predecessor tasks are complete.

Barriers can work with an arbitrary number of threads. This is a generalization of a Rendezvous which is defined for only two threads.

Implemented as a two-phase cyclic barrier, Barrier objects are suitable for use in loops. The separate filling and draining phases assure that all threads get released (drained) before any one of them can loop back and re-enter the barrier. The barrier fully resets after each cycle.

Example of using barriers:

from threading import Barrier, Thread

def get_votes(site):
    ballots = conduct_election(site)
    all_polls_closed.wait()        # do not count until all polls are closed
    totals = summarize(ballots)
    publish(site, totals)

all_polls_closed = Barrier(len(sites))
for site in sites:
    Thread(target=get_votes, args=(site,)).start()

In this example, the barrier enforces a rule that votes cannot be counted at any polling site until all polls are closed. Notice how a solution with a barrier is similar to one with threading.Thread.join(), but the threads stay alive and continue to do work (summarizing ballots) after the barrier point is crossed.

If any of the predecessor tasks can hang or be delayed, a barrier can be created with an optional timeout parameter. Then if the timeout period elapses before all the predecessor tasks reach the barrier point, all waiting threads are released and a BrokenBarrierError exception is raised:

def get_votes(site):
    ballots = conduct_election(site)
        all_polls_closed.wait(timeout=midnight - time.now())
    except BrokenBarrierError:
        lockbox = seal_ballots(ballots)
        totals = summarize(ballots)
        publish(site, totals)

In this example, the barrier enforces a more robust rule. If some election sites do not finish before midnight, the barrier times-out and the ballots are sealed and deposited in a queue for later handling.

See Barrier Synchronization Patterns for more examples of how barriers can be used in parallel computing. Also, there is a simple but thorough explanation of barriers in The Little Book of Semaphores , section 3.6.

(Contributed by Kristján Valur Jónsson with an API review by Jeffrey Yasskin in bpo-8777 .)

datetime and time

  • The datetime module has a new type timezone that implements the tzinfo interface by returning a fixed UTC offset and timezone name. This makes it easier to create timezone-aware datetime objects:

    >>> from datetime import datetime, timezone
    >>> datetime.now(timezone.utc)
    datetime.datetime(2010, 12, 8, 21, 4, 2, 923754, tzinfo=datetime.timezone.utc)
    >>> datetime.strptime("01/01/2000 12:00 +0000", "%m/%d/%Y %H:%M %z")
    datetime.datetime(2000, 1, 1, 12, 0, tzinfo=datetime.timezone.utc)
  • Also, timedelta objects can now be multiplied by float and divided by float and int objects. And timedelta objects can now divide one another.

  • The datetime.date.strftime() method is no longer restricted to years after 1900. The new supported year range is from 1000 to 9999 inclusive.

  • Whenever a two-digit year is used in a time tuple, the interpretation has been governed by time.accept2dyear. The default is True which means that for a two-digit year, the century is guessed according to the POSIX rules governing the %y strptime format.

    Starting with Py3.2, use of the century guessing heuristic will emit a DeprecationWarning. Instead, it is recommended that time.accept2dyear be set to False so that large date ranges can be used without guesswork:

    >>> import time, warnings
    >>> warnings.resetwarnings()      # remove the default warning filters
    >>> time.accept2dyear = True      # guess whether 11 means 11 or 2011
    >>> time.asctime((11, 1, 1, 12, 34, 56, 4, 1, 0))
    Warning (from warnings module):
    DeprecationWarning: Century info guessed for a 2-digit year.
    'Fri Jan  1 12:34:56 2011'
    >>> time.accept2dyear = False     # use the full range of allowable dates
    >>> time.asctime((11, 1, 1, 12, 34, 56, 4, 1, 0))
    'Fri Jan  1 12:34:56 11'

    Several functions now have significantly expanded date ranges. When time.accept2dyear is false, the time.asctime() function will accept any year that fits in a C int, while the time.mktime() and time.strftime() functions will accept the full range supported by the corresponding operating system functions.

(Contributed by Alexander Belopolsky and Victor Stinner in bpo-1289118 , bpo-5094 , bpo-6641 , bpo-2706 , bpo-1777412 , bpo-8013 , and bpo-10827 .)


The math module has been updated with six new functions inspired by the C99 standard.

The isfinite() function provides a reliable and fast way to detect special values. It returns True for regular numbers and False for Nan or Infinity:

>>> from math import isfinite
>>> [isfinite(x) for x in (123, 4.56, float('Nan'), float('Inf'))]
[True, True, False, False]

The expm1() function computes e**x-1 for small values of x without incurring the loss of precision that usually accompanies the subtraction of nearly equal quantities:

>>> from math import expm1
>>> expm1(0.013671875)   # more accurate way to compute e**x-1 for a small x

The erf() function computes a probability integral or Gaussian error function . The complementary error function, erfc(), is 1 - erf(x):

>>> from math import erf, erfc, sqrt
>>> erf(1.0/sqrt(2.0))   # portion of normal distribution within 1 standard deviation
>>> erfc(1.0/sqrt(2.0))  # portion of normal distribution outside 1 standard deviation
>>> erf(1.0/sqrt(2.0)) + erfc(1.0/sqrt(2.0))

The gamma() function is a continuous extension of the factorial function. See https://en.wikipedia.org/wiki/Gamma_function for details. Because the function is related to factorials, it grows large even for small values of x, so there is also a lgamma() function for computing the natural logarithm of the gamma function:

>>> from math import gamma, lgamma
>>> gamma(7.0)           # six factorial
>>> lgamma(801.0)        # log(800 factorial)

(Contributed by Mark Dickinson.)


The abc module now supports abstractclassmethod() and abstractstaticmethod().

These tools make it possible to define an abstract base class that requires a particular classmethod() or staticmethod() to be implemented:

class Temperature(metaclass=abc.ABCMeta):
    def from_fahrenheit(cls, t):
    def from_celsius(cls, t):

(Patch submitted by Daniel Urban; bpo-5867 .)


The io.BytesIO has a new method, getbuffer(), which provides functionality similar to memoryview(). It creates an editable view of the data without making a copy. The buffer’s random access and support for slice notation are well-suited to in-place editing:

>>> REC_LEN, LOC_START, LOC_LEN = 34, 7, 11

>>> def change_location(buffer, record_number, location):
...     start = record_number * REC_LEN + LOC_START
...     buffer[start: start+LOC_LEN] = location

>>> import io

>>> byte_stream = io.BytesIO(
...     b'G3805  storeroom  Main chassis    '
...     b'X7899  shipping   Reserve cog     '
...     b'L6988  receiving  Primary sprocket'
... )
>>> buffer = byte_stream.getbuffer()
>>> change_location(buffer, 1, b'warehouse  ')
>>> change_location(buffer, 0, b'showroom   ')
>>> print(byte_stream.getvalue())
b'G3805  showroom   Main chassis    '
b'X7899  warehouse  Reserve cog     '
b'L6988  receiving  Primary sprocket'

(Contributed by Antoine Pitrou in bpo-5506 .)


When writing a __repr__() method for a custom container, it is easy to forget to handle the case where a member refers back to the container itself. Python’s builtin objects such as list and set handle self-reference by displaying “…” in the recursive part of the representation string.

To help write such __repr__() methods, the reprlib module has a new decorator, recursive_repr(), for detecting recursive calls to __repr__() and substituting a placeholder string instead:

>>> class MyList(list):
...     @recursive_repr()
...     def __repr__(self):
...         return '<' + '|'.join(map(repr, self)) + '>'
>>> m = MyList('abc')
>>> m.append(m)
>>> m.append('x')
>>> print(m)

(Contributed by Raymond Hettinger in bpo-9826 and bpo-9840 .)


In addition to dictionary-based configuration described above, the logging package has many other improvements.

The logging documentation has been augmented by a basic tutorial, an advanced tutorial, and a cookbook of logging recipes. These documents are the fastest way to learn about logging.

The logging.basicConfig() set-up function gained a style argument to support three different types of string formatting. It defaults to “%” for traditional %-formatting, can be set to “{” for the new str.format() style, or can be set to “$” for the shell-style formatting provided by string.Template. The following three configurations are equivalent:

>>> from logging import basicConfig
>>> basicConfig(style='%', format="%(name)s -> %(levelname)s: %(message)s")
>>> basicConfig(style='{', format="{name} -> {levelname} {message}")
>>> basicConfig(style='$', format="$name -> $levelname: $message")

If no configuration is set-up before a logging event occurs, there is now a default configuration using a StreamHandler directed to sys.stderr for events of WARNING level or higher. Formerly, an event occurring before a configuration was set-up would either raise an exception or silently drop the event depending on the value of logging.raiseExceptions. The new default handler is stored in logging.lastResort.

The use of filters has been simplified. Instead of creating a Filter object, the predicate can be any Python callable that returns True or False.

There were a number of other improvements that add flexibility and simplify configuration. See the module documentation for a full listing of changes in Python 3.2.


The csv module now supports a new dialect, unix_dialect, which applies quoting for all fields and a traditional Unix style with '\n' as the line terminator. The registered dialect name is unix.

The csv.DictWriter has a new method, writeheader() for writing-out an initial row to document the field names:

>>> import csv, sys
>>> w = csv.DictWriter(sys.stdout, ['name', 'dept'], dialect='unix')
>>> w.writeheader()
>>> w.writerows([
...     {'name': 'tom', 'dept': 'accounting'},
...     {'name': 'susan', 'dept': 'Salesl'}])

(New dialect suggested by Jay Talbot in bpo-5975 , and the new method suggested by Ed Abraham in bpo-1537721 .)


There is a new and slightly mind-blowing tool ContextDecorator that is helpful for creating a context manager that does double duty as a function decorator.

As a convenience, this new functionality is used by contextmanager() so that no extra effort is needed to support both roles.

The basic idea is that both context managers and function decorators can be used for pre-action and post-action wrappers. Context managers wrap a group of statements using a with statement, and function decorators wrap a group of statements enclosed in a function. So, occasionally there is a need to write a pre-action or post-action wrapper that can be used in either role.

For example, it is sometimes useful to wrap functions or groups of statements with a logger that can track the time of entry and time of exit. Rather than writing both a function decorator and a context manager for the task, the contextmanager() provides both capabilities in a single definition:

from contextlib import contextmanager
import logging


def track_entry_and_exit(name):
    logging.info('Entering: %s', name)
    logging.info('Exiting: %s', name)

Formerly, this would have only been usable as a context manager:

with track_entry_and_exit('widget loader'):
    print('Some time consuming activity goes here')

Now, it can be used as a decorator as well:

@track_entry_and_exit('widget loader')
def activity():
    print('Some time consuming activity goes here')

Trying to fulfill two roles at once places some limitations on the technique. Context managers normally have the flexibility to return an argument usable by a with statement, but there is no parallel for function decorators.

In the above example, there is not a clean way for the track_entry_and_exit context manager to return a logging instance for use in the body of enclosed statements.

(Contributed by Michael Foord in bpo-9110 .)

decimal and fractions

Mark Dickinson crafted an elegant and efficient scheme for assuring that different numeric datatypes will have the same hash value whenever their actual values are equal (bpo-8188 ):

assert hash(Fraction(3, 2)) == hash(1.5) == \
       hash(Decimal("1.5")) == hash(complex(1.5, 0))

Some of the hashing details are exposed through a new attribute, sys.hash_info, which describes the bit width of the hash value, the prime modulus, the hash values for infinity and nan, and the multiplier used for the imaginary part of a number:

>>> sys.hash_info 
sys.hash_info(width=64, modulus=2305843009213693951, inf=314159, nan=0, imag=1000003)

An early decision to limit the inter-operability of various numeric types has been relaxed. It is still unsupported (and ill-advised) to have implicit mixing in arithmetic expressions such as Decimal('1.1') + float('1.1') because the latter loses information in the process of constructing the binary float. However, since existing floating point value can be converted losslessly to either a decimal or rational representation, it makes sense to add them to the constructor and to support mixed-type comparisons.

Similar changes were made to fractions.Fraction so that the from_float() and from_decimal() methods are no longer needed (bpo-8294 ):

>>> from decimal import Decimal
>>> from fractions import Fraction
>>> Decimal(1.1)
>>> Fraction(1.1)
Fraction(2476979795053773, 2251799813685248)

Another useful change for the decimal module is that the Context.clamp attribute is now public. This is useful in creating contexts that correspond to the decimal interchange formats specified in IEEE 754 (see bpo-8540 ).

(Contributed by Mark Dickinson and Raymond Hettinger.)


The ftplib.FTP class now supports the context management protocol to unconditionally consume socket.error exceptions and to close the FTP connection when done:

>>> from ftplib import FTP
>>> with FTP("ftp1.at.proftpd.org") as ftp:

'230 Anonymous login ok, restrictions apply.'
dr-xr-xr-x   9 ftp      ftp           154 May  6 10:43 .
dr-xr-xr-x   9 ftp      ftp           154 May  6 10:43 ..
dr-xr-xr-x   5 ftp      ftp          4096 May  6 10:43 CentOS
dr-xr-xr-x   3 ftp      ftp            18 Jul 10  2008 Fedora

Other file-like objects such as mmap.mmap and fileinput.input() also grew auto-closing context managers:

with fileinput.input(files=('log1.txt', 'log2.txt')) as f:
    for line in f:

(Contributed by Tarek Ziadé and Giampaolo Rodolà in bpo-4972 , and by Georg Brandl in bpo-8046 and bpo-1286 .)

The FTP_TLS class now accepts a context parameter, which is a ssl.SSLContext object allowing bundling SSL configuration options, certificates and private keys into a single (potentially long-lived) structure.

(Contributed by Giampaolo Rodolà; bpo-8806 .)


The os.popen() and subprocess.Popen() functions now support with statements for auto-closing of the file descriptors.

(Contributed by Antoine Pitrou and Brian Curtin in bpo-7461 and bpo-10554 .)


The select module now exposes a new, constant attribute, PIPE_BUF, which gives the minimum number of bytes which are guaranteed not to block when select.select() says a pipe is ready for writing.

>>> import select
>>> select.PIPE_BUF  

(Available on Unix systems. Patch by Sébastien Sablé in bpo-9862 )

gzip and zipfile

gzip.GzipFile now implements the io.BufferedIOBase abstract base class (except for truncate()). It also has a peek() method and supports unseekable as well as zero-padded file objects.

The gzip module also gains the compress() and decompress() functions for easier in-memory compression and decompression. Keep in mind that text needs to be encoded as bytes before compressing and decompressing:

>>> import gzip
>>> s = 'Three shall be the number thou shalt count, '
>>> s += 'and the number of the counting shall be three'
>>> b = s.encode()                        # convert to utf-8
>>> len(b)
>>> c = gzip.compress(b)
>>> len(c)
>>> gzip.decompress(c).decode()[:42]      # decompress and convert to text
'Three shall be the number thou shalt count'

(Contributed by Anand B. Pillai in bpo-3488 ; and by Antoine Pitrou, Nir Aides and Brian Curtin in bpo-9962 , bpo-1675951 , bpo-7471 and bpo-2846 .)

Also, the zipfile.ZipExtFile class was reworked internally to represent files stored inside an archive. The new implementation is significantly faster and can be wrapped in an io.BufferedReader object for more speedups. It also solves an issue where interleaved calls to read and readline gave the wrong results.

(Patch submitted by Nir Aides in bpo-7610 .)


The TarFile class can now be used as a context manager. In addition, its add() method has a new option, filter, that controls which files are added to the archive and allows the file metadata to be edited.

The new filter option replaces the older, less flexible exclude parameter which is now deprecated. If specified, the optional filter parameter needs to be a keyword argument. The user-supplied filter function accepts a TarInfo object and returns an updated TarInfo object, or if it wants the file to be excluded, the function can return None:

>>> import tarfile, glob

>>> def myfilter(tarinfo):
...     if tarinfo.isfile():             # only save real files
...         tarinfo.uname = 'monty'      # redact the user name
...         return tarinfo

>>> with tarfile.open(name='myarchive.tar.gz', mode='w:gz') as tf:
...     for filename in glob.glob('*.txt'):
...         tf.add(filename, filter=myfilter)
...     tf.list()
-rw-r--r-- monty/501        902 2011-01-26 17:59:11 annotations.txt
-rw-r--r-- monty/501        123 2011-01-26 17:59:11 general_questions.txt
-rw-r--r-- monty/501       3514 2011-01-26 17:59:11 prion.txt
-rw-r--r-- monty/501        124 2011-01-26 17:59:11 py_todo.txt
-rw-r--r-- monty/501       1399 2011-01-26 17:59:11 semaphore_notes.txt

(Proposed by Tarek Ziadé and implemented by Lars Gustäbel in bpo-6856 .)


The hashlib module has two new constant attributes listing the hashing algorithms guaranteed to be present in all implementations and those available on the current implementation:

>>> import hashlib

>>> hashlib.algorithms_guaranteed
{'sha1', 'sha224', 'sha384', 'sha256', 'sha512', 'md5'}

>>> hashlib.algorithms_available
{'md2', 'SHA256', 'SHA512', 'dsaWithSHA', 'mdc2', 'SHA224', 'MD4', 'sha256',
'sha512', 'ripemd160', 'SHA1', 'MDC2', 'SHA', 'SHA384', 'MD2',
'ecdsa-with-SHA1','md4', 'md5', 'sha1', 'DSA-SHA', 'sha224',
'dsaEncryption', 'DSA', 'RIPEMD160', 'sha', 'MD5', 'sha384'}

(Suggested by Carl Chenet in bpo-7418 .)


The ast module has a wonderful a general-purpose tool for safely evaluating expression strings using the Python literal syntax. The ast.literal_eval() function serves as a secure alternative to the builtin eval() function which is easily abused. Python 3.2 adds bytes and set literals to the list of supported types: strings, bytes, numbers, tuples, lists, dicts, sets, booleans, and None.

>>> from ast import literal_eval

>>> request = "{'req': 3, 'func': 'pow', 'args': (2, 0.5)}"
>>> literal_eval(request)
{'args': (2, 0.5), 'req': 3, 'func': 'pow'}

>>> request = "os.system('do something harmful')"
>>> literal_eval(request)
Traceback (most recent call last):
ValueError: malformed node or string: <_ast.Call object at 0x101739a10>

(Implemented by Benjamin Peterson and Georg Brandl.)


Different operating systems use various encodings for filenames and environment variables. The os module provides two new functions, fsencode() and fsdecode(), for encoding and decoding filenames:

>>> import os
>>> filename = 'Sehenswürdigkeiten'
>>> os.fsencode(filename)

Some operating systems allow direct access to encoded bytes in the environment. If so, the os.supports_bytes_environ constant will be true.

For direct access to encoded environment variables (if available), use the new os.getenvb() function or use os.environb which is a bytes version of os.environ.

(Contributed by Victor Stinner.)


The shutil.copytree() function has two new options:

  • ignore_dangling_symlinks: when symlinks=False so that the function copies a file pointed to by a symlink, not the symlink itself. This option will silence the error raised if the file doesn’t exist.

  • copy_function: is a callable that will be used to copy files. shutil.copy2() is used by default.

(Contributed by Tarek Ziadé.)

In addition, the shutil module now supports archiving operations for zipfiles, uncompressed tarfiles, gzipped tarfiles, and bzipped tarfiles. And there are functions for registering additional archiving file formats (such as xz compressed tarfiles or custom formats).

The principal functions are make_archive() and unpack_archive(). By default, both operate on the current directory (which can be set by os.chdir()) and on any sub-directories. The archive filename needs to be specified with a full pathname. The archiving step is non-destructive (the original files are left unchanged).

>>> import shutil, pprint

>>> os.chdir('mydata')  # change to the source directory
>>> f = shutil.make_archive('/var/backup/mydata',
...                         'zip')      # archive the current directory
>>> f                                   # show the name of archive
>>> os.chdir('tmp')                     # change to an unpacking
>>> shutil.unpack_archive('/var/backup/mydata.zip')  # recover the data

>>> pprint.pprint(shutil.get_archive_formats())  # display known formats
[('bztar', "bzip2'ed tar-file"),
 ('gztar', "gzip'ed tar-file"),
 ('tar', 'uncompressed tar file'),
 ('zip', 'ZIP file')]

>>> shutil.register_archive_format(     # register a new archive format
...     name='xz',
...     function=xz.compress,           # callable archiving function
...     extra_args=[('level', 8)],      # arguments to the function
...     description='xz compression'
... )

(Contributed by Tarek Ziadé.)


The sqlite3 module was updated to pysqlite version 2.6.0. It has two new capabilities.

(Contributed by R. David Murray and Shashwat Anand; bpo-8845 .)


A new html module was introduced with only a single function, escape(), which is used for escaping reserved characters from HTML markup:

>>> import html
>>> html.escape('x > 2 && x < 7')
'x &gt; 2 &amp;&amp; x &lt; 7'


The socket module has two new improvements.

  • Socket objects now have a detach() method which puts the socket into closed state without actually closing the underlying file descriptor. The latter can then be reused for other purposes. (Added by Antoine Pitrou; bpo-8524 .)

  • socket.create_connection() now supports the context management protocol to unconditionally consume socket.error exceptions and to close the socket when done. (Contributed by Giampaolo Rodolà; bpo-9794 .)


The ssl module added a number of features to satisfy common requirements for secure (encrypted, authenticated) internet connections:

  • A new class, SSLContext, serves as a container for persistent SSL data, such as protocol settings, certificates, private keys, and various other options. It includes a wrap_socket() for creating an SSL socket from an SSL context.

  • A new function, ssl.match_hostname(), supports server identity verification for higher-level protocols by implementing the rules of HTTPS (from RFC 2818 ) which are also suitable for other protocols.

  • The ssl.wrap_socket() constructor function now takes a ciphers argument. The ciphers string lists the allowed encryption algorithms using the format described in the OpenSSL documentation .

  • When linked against recent versions of OpenSSL, the ssl module now supports the Server Name Indication extension to the TLS protocol, allowing multiple “virtual hosts” using different certificates on a single IP port. This extension is only supported in client mode, and is activated by passing the server_hostname argument to