What’s New In Python 3.3

This article explains the new features in Python 3.3, compared to 3.2. Python 3.3 was released on September 29, 2012. For full details, see the changelog .

See also

PEP 398 - Python 3.3 Release Schedule

Summary – Release highlights

New syntax features:

New library modules:

  • faulthandler (helps debugging low-level crashes)

  • ipaddress (high-level objects representing IP addresses and masks)

  • lzma (compress data using the XZ / LZMA algorithm)

  • unittest.mock (replace parts of your system under test with mock objects)

  • venv (Python virtual environments, as in the popular virtualenv package)

New built-in features:

Implementation improvements:

Significantly Improved Library Modules:

Security improvements:

  • Hash randomization is switched on by default.

Please read on for a comprehensive list of user-facing changes.

PEP 405: Virtual Environments

Virtual environments help create separate Python setups while sharing a system-wide base install, for ease of maintenance. Virtual environments have their own set of private site packages (i.e. locally-installed libraries), and are optionally segregated from the system-wide site packages. Their concept and implementation are inspired by the popular virtualenv third-party package, but benefit from tighter integration with the interpreter core.

This PEP adds the venv module for programmatic access, and the pyvenv script for command-line access and administration. The Python interpreter checks for a pyvenv.cfg, file whose existence signals the base of a virtual environment’s directory tree.

See also

PEP 405 - Python Virtual Environments

PEP written by Carl Meyer; implementation by Carl Meyer and Vinay Sajip

PEP 420: Implicit Namespace Packages

Native support for package directories that don’t require __init__.py marker files and can automatically span multiple path segments (inspired by various third party approaches to namespace packages, as described in PEP 420 )

See also

PEP 420 - Implicit Namespace Packages

PEP written by Eric V. Smith; implementation by Eric V. Smith and Barry Warsaw

PEP 3118: New memoryview implementation and buffer protocol documentation

The implementation of PEP 3118 has been significantly improved.

The new memoryview implementation comprehensively fixes all ownership and lifetime issues of dynamically allocated fields in the Py_buffer struct that led to multiple crash reports. Additionally, several functions that crashed or returned incorrect results for non-contiguous or multi-dimensional input have been fixed.

The memoryview object now has a PEP-3118 compliant getbufferproc() that checks the consumer’s request type. Many new features have been added, most of them work in full generality for non-contiguous arrays and arrays with suboffsets.

The documentation has been updated, clearly spelling out responsibilities for both exporters and consumers. Buffer request flags are grouped into basic and compound flags. The memory layout of non-contiguous and multi-dimensional NumPy-style arrays is explained.


  • All native single character format specifiers in struct module syntax (optionally prefixed with ‘@’) are now supported.

  • With some restrictions, the cast() method allows changing of format and shape of C-contiguous arrays.

  • Multi-dimensional list representations are supported for any array type.

  • Multi-dimensional comparisons are supported for any array type.

  • One-dimensional memoryviews of hashable (read-only) types with formats B, b or c are now hashable. (Contributed by Antoine Pitrou in bpo-13411 .)

  • Arbitrary slicing of any 1-D arrays type is supported. For example, it is now possible to reverse a memoryview in O(1) by using a negative step.

API changes

  • The maximum number of dimensions is officially limited to 64.

  • The representation of empty shape, strides and suboffsets is now an empty tuple instead of None.

  • Accessing a memoryview element with format ‘B’ (unsigned bytes) now returns an integer (in accordance with the struct module syntax). For returning a bytes object the view must be cast to ‘c’ first.

  • memoryview comparisons now use the logical structure of the operands and compare all array elements by value. All format strings in struct module syntax are supported. Views with unrecognised format strings are still permitted, but will always compare as unequal, regardless of view contents.

  • For further changes see Build and C API Changes and Porting C code.

(Contributed by Stefan Krah in bpo-10181 .)

See also

PEP 3118 - Revising the Buffer Protocol

PEP 393: Flexible String Representation

The Unicode string type is changed to support multiple internal representations, depending on the character with the largest Unicode ordinal (1, 2, or 4 bytes) in the represented string. This allows a space-efficient representation in common cases, but gives access to full UCS-4 on all systems. For compatibility with existing APIs, several representations may exist in parallel; over time, this compatibility should be phased out.

On the Python side, there should be no downside to this change.

On the C API side, PEP 393 is fully backward compatible. The legacy API should remain available at least five years. Applications using the legacy API will not fully benefit of the memory reduction, or - worse - may use a bit more memory, because Python may have to maintain two versions of each string (in the legacy format and in the new efficient storage).


Changes introduced by PEP 393 are the following:

  • Python now always supports the full range of Unicode code points, including non-BMP ones (i.e. from U+0000 to U+10FFFF). The distinction between narrow and wide builds no longer exists and Python now behaves like a wide build, even under Windows.

  • With the death of narrow builds, the problems specific to narrow builds have also been fixed, for example:

    • len() now always returns 1 for non-BMP characters, so len('\U0010FFFF') == 1;

    • surrogate pairs are not recombined in string literals, so '\uDBFF\uDFFF' != '\U0010FFFF';

    • indexing or slicing non-BMP characters returns the expected value, so '\U0010FFFF'[0] now returns '\U0010FFFF' and not '\uDBFF';

    • all other functions in the standard library now correctly handle non-BMP code points.

  • The value of sys.maxunicode is now always 1114111 (0x10FFFF in hexadecimal). The PyUnicode_GetMax() function still returns either 0xFFFF or 0x10FFFF for backward compatibility, and it should not be used with the new Unicode API (see bpo-13054 ).

  • The ./configure flag --with-wide-unicode has been removed.

Performance and resource usage

The storage of Unicode strings now depends on the highest code point in the string:

  • pure ASCII and Latin1 strings (U+0000-U+00FF) use 1 byte per code point;

  • BMP strings (U+0000-U+FFFF) use 2 bytes per code point;

  • non-BMP strings (U+10000-U+10FFFF) use 4 bytes per code point.

The net effect is that for most applications, memory usage of string storage should decrease significantly - especially compared to former wide unicode builds - as, in many cases, strings will be pure ASCII even in international contexts (because many strings store non-human language data, such as XML fragments, HTTP headers, JSON-encoded data, etc.). We also hope that it will, for the same reasons, increase CPU cache efficiency on non-trivial applications. The memory usage of Python 3.3 is two to three times smaller than Python 3.2, and a little bit better than Python 2.7, on a Django benchmark (see the PEP for details).

See also

PEP 393 - Flexible String Representation

PEP written by Martin von Löwis; implementation by Torsten Becker and Martin von Löwis.

PEP 397: Python Launcher for Windows

The Python 3.3 Windows installer now includes a py launcher application that can be used to launch Python applications in a version independent fashion.

This launcher is invoked implicitly when double-clicking *.py files. If only a single Python version is installed on the system, that version will be used to run the file. If multiple versions are installed, the most recent version is used by default, but this can be overridden by including a Unix-style “shebang line” in the Python script.

The launcher can also be used explicitly from the command line as the py application. Running py follows the same version selection rules as implicitly launching scripts, but a more specific version can be selected by passing appropriate arguments (such as -3 to request Python 3 when Python 2 is also installed, or -2.6 to specifically request an earlier Python version when a more recent version is installed).

In addition to the launcher, the Windows installer now includes an option to add the newly installed Python to the system PATH. (Contributed by Brian Curtin in bpo-3561 .)

See also

PEP 397 - Python Launcher for Windows

PEP written by Mark Hammond and Martin v. Löwis; implementation by Vinay Sajip.

Launcher documentation: Python Launcher for Windows

Installer PATH modification: Finding the Python executable

PEP 3151: Reworking the OS and IO exception hierarchy

The hierarchy of exceptions raised by operating system errors is now both simplified and finer-grained.

You don’t have to worry anymore about choosing the appropriate exception type between OSError, IOError, EnvironmentError, WindowsError, mmap.error, socket.error or select.error. All these exception types are now only one: OSError. The other names are kept as aliases for compatibility reasons.

Also, it is now easier to catch a specific error condition. Instead of inspecting the errno attribute (or args[0]) for a particular constant from the errno module, you can catch the adequate OSError subclass. The available subclasses are the following:

And the ConnectionError itself has finer-grained subclasses:

Thanks to the new exceptions, common usages of the errno can now be avoided. For example, the following code written for Python 3.2:

from errno import ENOENT, EACCES, EPERM

    with open("document.txt") as f:
        content = f.read()
except IOError as err:
    if err.errno == ENOENT:
        print("document.txt file is missing")
    elif err.errno in (EACCES, EPERM):
        print("You are not allowed to read document.txt")

can now be written without the errno import and without manual inspection of exception attributes:

    with open("document.txt") as f:
        content = f.read()
except FileNotFoundError:
    print("document.txt file is missing")
except PermissionError:
    print("You are not allowed to read document.txt")

See also

PEP 3151 - Reworking the OS and IO Exception Hierarchy

PEP written and implemented by Antoine Pitrou

PEP 380: Syntax for Delegating to a Subgenerator

PEP 380 adds the yield from expression, allowing a generator to delegate part of its operations to another generator. This allows a section of code containing yield to be factored out and placed in another generator. Additionally, the subgenerator is allowed to return with a value, and the value is made available to the delegating generator.

While designed primarily for use in delegating to a subgenerator, the yield from expression actually allows delegation to arbitrary subiterators.

For simple iterators, yield from iterable is essentially just a shortened form of for item in iterable: yield item:

>>> def g(x):
...     yield from range(x, 0, -1)
...     yield from range(x)
>>> list(g(5))
[5, 4, 3, 2, 1, 0, 1, 2, 3, 4]

However, unlike an ordinary loop, yield from allows subgenerators to receive sent and thrown values directly from the calling scope, and return a final value to the outer generator:

>>> def accumulate():
...     tally = 0
...     while 1:
...         next = yield
...         if next is None:
...             return tally
...         tally += next
>>> def gather_tallies(tallies):
...     while 1:
...         tally = yield from accumulate()
...         tallies.append(tally)
>>> tallies = []
>>> acc = gather_tallies(tallies)
>>> next(acc)  # Ensure the accumulator is ready to accept values
>>> for i in range(4):
...     acc.send(i)
>>> acc.send(None)  # Finish the first tally
>>> for i in range(5):
...     acc.send(i)
>>> acc.send(None)  # Finish the second tally
>>> tallies
[6, 10]

The main principle driving this change is to allow even generators that are designed to be used with the send and throw methods to be split into multiple subgenerators as easily as a single large function can be split into multiple subfunctions.

See also

PEP 380 - Syntax for Delegating to a Subgenerator

PEP written by Greg Ewing; implementation by Greg Ewing, integrated into 3.3 by Renaud Blanch, Ryan Kelly and Nick Coghlan; documentation by Zbigniew Jędrzejewski-Szmek and Nick Coghlan

PEP 409: Suppressing exception context

PEP 409 introduces new syntax that allows the display of the chained exception context to be disabled. This allows cleaner error messages in applications that convert between exception types:

>>> class D:
...     def __init__(self, extra):
...         self._extra_attributes = extra
...     def __getattr__(self, attr):
...         try:
...             return self._extra_attributes[attr]
...         except KeyError:
...             raise AttributeError(attr) from None
>>> D({}).x
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 8, in __getattr__
AttributeError: x

Without the from None suffix to suppress the cause, the original exception would be displayed by default:

>>> class C:
...     def __init__(self, extra):
...         self._extra_attributes = extra
...     def __getattr__(self, attr):
...         try:
...             return self._extra_attributes[attr]
...         except KeyError:
...             raise AttributeError(attr)
>>> C({}).x
Traceback (most recent call last):
  File "<stdin>", line 6, in __getattr__
KeyError: 'x'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 8, in __getattr__
AttributeError: x

No debugging capability is lost, as the original exception context remains available if needed (for example, if an intervening library has incorrectly suppressed valuable underlying details):

>>> try:
...     D({}).x
... except AttributeError as exc:
...     print(repr(exc.__context__))

See also

PEP 409 - Suppressing exception context

PEP written by Ethan Furman; implemented by Ethan Furman and Nick Coghlan.

PEP 414: Explicit Unicode literals

To ease the transition from Python 2 for Unicode aware Python applications that make heavy use of Unicode literals, Python 3.3 once again supports the “u” prefix for string literals. This prefix has no semantic significance in Python 3, it is provided solely to reduce the number of purely mechanical changes in migrating to Python 3, making it easier for developers to focus on the more significant semantic changes (such as the stricter default separation of binary and text data).

See also

PEP 414 - Explicit Unicode literals

PEP written by Armin Ronacher.

PEP 3155: Qualified name for classes and functions

Functions and class objects have a new __qualname__ attribute representing the “path” from the module top-level to their definition. For global functions and classes, this is the same as __name__. For other functions and classes, it provides better information about where they were actually defined, and how they might be accessible from the global scope.

Example with (non-bound) methods:

>>> class C:
...     def meth(self):
...         pass
>>> C.meth.__name__
>>> C.meth.__qualname__

Example with nested classes:

>>> class C:
...     class D:
...         def meth(self):
...             pass
>>> C.D.__name__
>>> C.D.__qualname__
>>> C.D.meth.__name__
>>> C.D.meth.__qualname__

Example with nested functions:

>>> def outer():
...     def inner():
...         pass
...     return inner
>>> outer().__name__
>>> outer().__qualname__

The string representation of those objects is also changed to include the new, more precise information:

>>> str(C.D)
"<class '__main__.C.D'>"
>>> str(C.D.meth)
'<function C.D.meth at 0x7f46b9fe31e0>'

See also

PEP 3155 - Qualified name for classes and functions

PEP written and implemented by Antoine Pitrou.

PEP 412: Key-Sharing Dictionary

Dictionaries used for the storage of objects’ attributes are now able to share part of their internal storage between each other (namely, the part which stores the keys and their respective hashes). This reduces the memory consumption of programs creating many instances of non-builtin types.

See also

PEP 412 - Key-Sharing Dictionary

PEP written and implemented by Mark Shannon.

PEP 362: Function Signature Object

A new function inspect.signature() makes introspection of python callables easy and straightforward. A broad range of callables is supported: python functions, decorated or not, classes, and functools.partial() objects. New classes inspect.Signature, inspect.Parameter and inspect.BoundArguments hold information about the call signatures, such as, annotations, default values, parameters kinds, and bound arguments, which considerably simplifies writing decorators and any code that validates or amends calling signatures or arguments.

See also

PEP 362 : - Function Signature Object

PEP written by Brett Cannon, Yury Selivanov, Larry Hastings, Jiwon Seo; implemented by Yury Selivanov.

PEP 421: Adding sys.implementation

A new attribute on the sys module exposes details specific to the implementation of the currently running interpreter. The initial set of attributes on sys.implementation are name, version, hexversion, and cache_tag.

The intention of sys.implementation is to consolidate into one namespace the implementation-specific data used by the standard library. This allows different Python implementations to share a single standard library code base much more easily. In its initial state, sys.implementation holds only a small portion of the implementation-specific data. Over time that ratio will shift in order to make the standard library more portable.

One example of improved standard library portability is cache_tag. As of Python 3.3, sys.implementation.cache_tag is used by importlib to support PEP 3147 compliance. Any Python implementation that uses importlib for its built-in import system may use cache_tag to control the caching behavior for modules.


The implementation of sys.implementation also introduces a new type to Python: types.SimpleNamespace. In contrast to a mapping-based namespace, like dict, SimpleNamespace is attribute-based, like object. However, unlike object, SimpleNamespace instances are writable. This means that you can add, remove, and modify the namespace through normal attribute access.

See also

PEP 421 - Adding sys.implementation

PEP written and implemented by Eric Snow.

Using importlib as the Implementation of Import

bpo-2377 - Replace __import__ w/ importlib.__import__ bpo-13959 - Re-implement parts of imp in pure Python bpo-14605 - Make import machinery explicit bpo-14646 - Require loaders set __loader__ and __package__

The __import__() function is now powered by importlib.__import__(). This work leads to the completion of “phase 2” of PEP 302 . There are multiple benefits to this change. First, it has allowed for more of the machinery powering import to be exposed instead of being implicit and hidden within the C code. It also provides a single implementation for all Python VMs supporting Python 3.3 to use, helping to end any VM-specific deviations in import semantics. And finally it eases the maintenance of import, allowing for future growth to occur.

For the common user, there should be no visible change in semantics. For those whose code currently manipulates import or calls import programmatically, the code changes that might possibly be required are covered in the Porting Python code section of this document.

New APIs

One of the large benefits of this work is the exposure of what goes into making the import statement work. That means the various importers that were once implicit are now fully exposed as part of the importlib package.

The abstract base classes defined in importlib.abc have been expanded to properly delineate between meta path finders and path entry finders by introducing importlib.abc.MetaPathFinder and importlib.abc.PathEntryFinder, respectively. The old ABC of importlib.abc.Finder is now only provided for backwards-compatibility and does not enforce any method requirements.

In terms of finders, importlib.machinery.FileFinder exposes the mechanism used to search for source and bytecode files of a module. Previously this class was an implicit member of sys.path_hooks.

For loaders, the new abstract base class importlib.abc.FileLoader helps write a loader that uses the file system as the storage mechanism for a module’s code. The loader for source files (importlib.machinery.SourceFileLoader), sourceless bytecode files (importlib.machinery.SourcelessFileLoader), and extension modules (importlib.machinery.ExtensionFileLoader) are now available for direct use.

ImportError now has name and path attributes which are set when there is relevant data to provide. The message for failed imports will also provide the full name of the module now instead of just the tail end of the module’s name.

The importlib.invalidate_caches() function will now call the method with the same name on all finders cached in sys.path_importer_cache to help clean up any stored state as necessary.

Visible Changes

For potential required changes to code, see the Porting Python code section.

Beyond the expanse of what importlib now exposes, there are other visible changes to import. The biggest is that sys.meta_path and sys.path_hooks now store all of the meta path finders and path entry hooks used by import. Previously the finders were implicit and hidden within the C code of import instead of being directly exposed. This means that one can now easily remove or change the order of the various finders to fit one’s needs.

Another change is that all modules have a __loader__ attribute, storing the loader used to create the module. PEP 302 has been updated to make this attribute mandatory for loaders to implement, so in the future once 3rd-party loaders have been updated people will be able to rely on the existence of the attribute. Until such time, though, import is setting the module post-load.

Loaders are also now expected to set the __package__ attribute from PEP 366 . Once again, import itself is already setting this on all loaders from importlib and import itself is setting the attribute post-load.

None is now inserted into sys.path_importer_cache when no finder can be found on sys.path_hooks. Since imp.NullImporter is not directly exposed on sys.path_hooks it could no longer be relied upon to always be available to use as a value representing no finder found.

All other changes relate to semantic changes which should be taken into consideration when updating code for Python 3.3, and thus should be read about in the Porting Python code section of this document.

(Implementation by Brett Cannon)

Other Language Changes

Some smaller changes made to the core Python language are:

  • Added support for Unicode name aliases and named sequences. Both unicodedata.lookup() and '\N{...}' now resolve name aliases, and unicodedata.lookup() resolves named sequences too.

    (Contributed by Ezio Melotti in bpo-12753 .)

  • Unicode database updated to UCD version 6.1.0

  • Equality comparisons on range() objects now return a result reflecting the equality of the underlying sequences generated by those range objects. (bpo-13201 )

  • The count(), find(), rfind(), index() and rindex() methods of bytes and bytearray objects now accept an integer between 0 and 255 as their first argument.

    (Contributed by Petri Lehtinen in bpo-12170 .)

  • The rjust(), ljust(), and center() methods of bytes and bytearray now accept a bytearray for the fill argument. (Contributed by Petri Lehtinen in bpo-12380 .)

  • New methods have been added to list and bytearray: copy() and clear() (bpo-10516 ). Consequently, MutableSequence now also defines a clear() method (bpo-11388 ).

  • Raw bytes literals can now be written rb"..." as well as br"...".

    (Contributed by Antoine Pitrou in bpo-13748 .)

  • dict.setdefault() now does only one lookup for the given key, making it atomic when used with built-in types.

    (Contributed by Filip Gruszczyński in bpo-13521 .)

  • The error messages produced when a function call does not match the function signature have been significantly improved.

    (Contributed by Benjamin Peterson.)

A Finer-Grained Import Lock

Previous versions of CPython have always relied on a global import lock. This led to unexpected annoyances, such as deadlocks when importing a module would trigger code execution in a different thread as a side-effect. Clumsy workarounds were sometimes employed, such as the PyImport_ImportModuleNoBlock() C API function.

In Python 3.3, importing a module takes a per-module lock. This correctly serializes importation of a given module from multiple threads (preventing the exposure of incompletely initialized modules), while eliminating the aforementioned annoyances.

(Contributed by Antoine Pitrou in bpo-9260 .)

Builtin functions and types

  • open() gets a new opener parameter: the underlying file descriptor for the file object is then obtained by calling opener with (file, flags). It can be used to use custom flags like os.O_CLOEXEC for example. The 'x' mode was added: open for exclusive creation, failing if the file already exists.

  • print(): added the flush keyword argument. If the flush keyword argument is true, the stream is forcibly flushed.

  • hash(): hash randomization is enabled by default, see object.__hash__() and PYTHONHASHSEED.

  • The str type gets a new casefold() method: return a casefolded copy of the string, casefolded strings may be used for caseless matching. For example, 'ß'.casefold() returns 'ss'.

  • The sequence documentation has been substantially rewritten to better explain the binary/text sequence distinction and to provide specific documentation sections for the individual builtin sequence types (bpo-4966 ).

New Modules


This new debug module faulthandler contains functions to dump Python tracebacks explicitly, on a fault (a crash like a segmentation fault), after a timeout, or on a user signal. Call faulthandler.enable() to install fault handlers for the SIGSEGV, SIGFPE, SIGABRT, SIGBUS, and SIGILL signals. You can also enable them at startup by setting the PYTHONFAULTHANDLER environment variable or by using -X faulthandler command line option.

Example of a segmentation fault on Linux:

$ python -q -X faulthandler
>>> import ctypes
>>> ctypes.string_at(0)
Fatal Python error: Segmentation fault

Current thread 0x00007fb899f39700:
  File "/home/python/cpython/Lib/ctypes/__init__.py", line 486 in string_at
  File "<stdin>", line 1 in <module>
Segmentation fault


The new ipaddress module provides tools for creating and manipulating objects representing IPv4 and IPv6 addresses, networks and interfaces (i.e. an IP address associated with a specific IP subnet).

(Contributed by Google and Peter Moody in PEP 3144 .)


The newly-added lzma module provides data compression and decompression using the LZMA algorithm, including support for the .xz and .lzma file formats.

(Contributed by Nadeem Vawda and Per Øyvind Karlsen in bpo-6715 .)

Improved Modules


Improved support for abstract base classes containing descriptors composed with abstract methods. The recommended approach to declaring abstract descriptors is now to provide __isabstractmethod__ as a dynamically updated property. The built-in descriptors have been updated accordingly.

(Contributed by Darren Dale in bpo-11610 .)

abc.ABCMeta.register() now returns the registered subclass, which means it can now be used as a class decorator (bpo-10868 ).


The array module supports the long long type using q and Q type codes.

(Contributed by Oren Tirosh and Hirokazu Yamamoto in bpo-1172711 .)


ASCII-only Unicode strings are now accepted by the decoding functions of the base64 modern interface. For example, base64.b64decode('YWJj') returns b'abc'. (Contributed by Catalin Iacob in bpo-13641 .)


In addition to the binary objects they normally accept, the a2b_ functions now all also accept ASCII-only strings as input. (Contributed by Antoine Pitrou in bpo-13637 .)


The bz2 module has been rewritten from scratch. In the process, several new features have been added:

  • New bz2.open() function: open a bzip2-compressed file in binary or text mode.

  • bz2.BZ2File can now read from and write to arbitrary file-like objects, by means of its constructor’s fileobj argument.

    (Contributed by Nadeem Vawda in bpo-5863 .)

  • bz2.BZ2File and bz2.decompress() can now decompress multi-stream inputs (such as those produced by the pbzip2 tool). bz2.BZ2File can now also be used to create this type of file, using the 'a' (append) mode.

    (Contributed by Nir Aides in bpo-1625 .)

  • bz2.BZ2File now implements all of the io.BufferedIOBase API, except for the detach() and truncate() methods.


The mbcs codec has been rewritten to handle correctly replace and ignore error handlers on all Windows versions. The mbcs codec now supports all error handlers, instead of only replace to encode and ignore to decode.

A new Windows-only codec has been added: cp65001 (bpo-13216 ). It is the Windows code page 65001 (Windows UTF-8, CP_UTF8). For example, it is used by sys.stdout if the console output code page is set to cp65001 (e.g., using chcp 65001 command).

Multibyte CJK decoders now resynchronize faster. They only ignore the first byte of an invalid byte sequence. For example, b'\xff\n'.decode('gb2312', 'replace') now returns a \n after the replacement character.

(bpo-12016 )

Incremental CJK codec encoders are no longer reset at each call to their encode() methods. For example:

>>> import codecs
>>> encoder = codecs.getincrementalencoder('hz')('strict')
>>> b''.join(encoder.encode(x) for x in '\u52ff\u65bd\u65bc\u4eba\u3002 Bye.')
b'~{NpJ)l6HK!#~} Bye.'

This example gives b'~{Np~}~{J)~}~{l6~}~{HK~}~{!#~} Bye.' with older Python versions.

(bpo-12100 )

The unicode_internal codec has been deprecated.


Addition of a new ChainMap class to allow treating a number of mappings as a single unit. (Written by Raymond Hettinger for bpo-11089 , made public in bpo-11297 .)

The abstract base classes have been moved in a new collections.abc module, to better differentiate between the abstract and the concrete collections classes. Aliases for ABCs are still present in the collections module to preserve existing imports. (bpo-11085 )

The Counter class now supports the unary + and - operators, as well as the in-place operators +=, -=, |=, and &=. (Contributed by Raymond Hettinger in bpo-13121 .)


ExitStack now provides a solid foundation for programmatic manipulation of context managers and similar cleanup functionality. Unlike the previous contextlib.nested API (which was deprecated and removed), the new API is designed to work correctly regardless of whether context managers acquire their resources in their __init__ method (for example, file objects) or in their __enter__ method (for example, synchronisation objects from the threading module).

(bpo-13585 )


Addition of salt and modular crypt format (hashing method) and the mksalt() function to the crypt module.

(bpo-10924 )


  • If the curses module is linked to the ncursesw library, use Unicode functions when Unicode strings or characters are passed (e.g. waddwstr()), and bytes functions otherwise (e.g. waddstr()).

  • Use the locale encoding instead of utf-8 to encode Unicode strings.

  • curses.window has a new curses.window.encoding attribute.

  • The curses.window class has a new get_wch() method to get a wide character

  • The curses module has a new unget_wch() function to push a wide character so the next get_wch() will return it

(Contributed by Iñigo Serna in bpo-6755 .)



bpo-7652 - integrate fast native decimal arithmetic.

C-module and libmpdec written by Stefan Krah.

The new C version of the decimal module integrates the high speed libmpdec library for arbitrary precision correctly-rounded decimal floating point arithmetic. libmpdec conforms to IBM’s General Decimal Arithmetic Specification.

Performance gains range from 10x for database applications to 100x for numerically intensive applications. These numbers are expected gains for standard precisions used in decimal floating point arithmetic. Since the precision is user configurable, the exact figures may vary. For example, in integer bignum arithmetic the differences can be significantly higher.

The following table is meant as an illustration. Benchmarks are available at http://www.bytereef.org/mpdecimal/quickstart.html .

















  • The FloatOperation signal optionally enables stricter semantics for mixing floats and Decimals.

  • If Python is compiled without threads, the C version automatically disables the expensive thread local context machinery. In this case, the variable HAVE_THREADS is set to False.

API changes

  • The C module has the following context limits, depending on the machine architecture:












  • In the context templates (DefaultContext, BasicContext and ExtendedContext) the magnitude of Emax and Emin has changed to 999999.

  • The Decimal constructor in decimal.py does not observe the context limits and converts values with arbitrary exponents or precision exactly. Since the C version has internal limits, the following scheme is used: If possible, values are converted exactly, otherwise InvalidOperation is raised and the result is NaN. In the latter case it is always possible to use create_decimal() in order to obtain a rounded or inexact value.

  • The power function in decimal.py is always correctly-rounded. In the C version, it is defined in terms of the correctly-rounded exp() and ln() functions, but the final result is only “almost always correctly rounded”.

  • In the C version, the context dictionary containing the signals is a MutableMapping. For speed reasons, flags and traps always refer to the same MutableMapping that the context was initialized with. If a new signal dictionary is assigned, flags and traps are updated with the new values, but they do not reference the RHS dictionary.

  • Pickling a Context produces a different output in order to have a common interchange format for the Python and C versions.

  • The order of arguments in the Context constructor has been changed to match the order displayed by repr().

  • The watchexp parameter in the quantize() method is deprecated.


Policy Framework

The email package now has a policy framework. A Policy is an object with several methods and properties that control how the email package behaves. The primary policy for Python 3.3 is the Compat32 policy, which provides backward compatibility with the email package in Python 3.2. A policy can be specified when an email message is parsed by a parser, or when a Message object is created, or when an email is serialized using a generator. Unless overridden, a policy passed to a parser is inherited by all the Message object and sub-objects created by the parser. By default a generator will use the policy of the Message object it is serializing. The default policy is compat32.

The minimum set of controls implemented by all policy objects are:


The maximum length, excluding the linesep character(s), individual lines may have when a Message is serialized. Defaults to 78.


The character used to separate individual lines when a Message is serialized. Defaults to \n.


7bit or 8bit. 8bit applies only to a Bytes generator, and means that non-ASCII may be used where allowed by the protocol (or where it exists in the original input).


Causes a parser to raise error when defects are encountered instead of adding them to the Message object’s defects list.

A new policy instance, with new settings, is created using the clone() method of policy objects. clone takes any of the above controls as keyword arguments. Any control not specified in the call retains its default value. Thus you can create a policy that uses \r\n linesep characters like this:

mypolicy = compat32.clone(linesep='\r\n')

Policies can be used to make the generation of messages in the format needed by your application simpler. Instead of having to remember to specify linesep='\r\n' in all the places you call a generator, you can specify it once, when you set the policy used by the parser or the Message, whichever your program uses to create Message objects. On the other hand, if you need to generate messages in multiple forms, you can still specify the parameters in the appropriate generator call. Or you can have custom policy instances for your different cases, and pass those in when you create the generator.

Provisional Policy with New Header API

While the policy framework is worthwhile all by itself, the main motivation for introducing it is to allow the creation of new policies that implement new features for the email package in a way that maintains backward compatibility for those who do not use the new policies. Because the new policies introduce a new API, we are releasing them in Python 3.3 as a provisional policy. Backwards incompatible changes (up to and including removal of the code) may occur if deemed necessary by the core developers.

The new policies are instances of EmailPolicy, and add the following additional controls:


Controls whether or not headers parsed by a parser are refolded by the generator. It can be none, long, or all. The default is long, which means that source headers with a line longer than max_line_length get refolded. none means no line get refolded, and all means that all lines get refolded.


A callable that take a name and value and produces a custom header object.

The header_factory is the key to the new features provided by the new policies. When one of the new policies is used, any header retrieved from a Message object is an object produced by the header_factory, and any time you set a header on a Message it becomes an object produced by header_factory. All such header objects have a name attribute equal to the header name. Address and Date headers have additional attributes that give you access to the parsed data of the header. This means you can now do things like this:

>>> m = Message(policy=SMTP)
>>> m['To'] = 'Éric <[email protected]>'
>>> m['to']
'Éric <[email protected]>'
>>> m['to'].addresses
(Address(display_name='Éric', username='foo', domain='example.com'),)
>>> m['to'].addresses[0].username
>>> m['to'].addresses[0].display_name
>>> m['Date'] = email.utils.localtime()
>>> m['Date'].datetime
datetime.datetime(2012, 5, 25, 21, 39, 24, 465484, tzinfo=datetime.timezone(datetime.timedelta(-1, 72000), 'EDT'))
>>> m['Date']
'Fri, 25 May 2012 21:44:27 -0400'
>>> print(m)
To: =?utf-8?q?=C3=89ric?= <[email protected]>
Date: Fri, 25 May 2012 21:44:27 -0400

You will note that the unicode display name is automatically encoded as utf-8 when the message is serialized, but that when the header is accessed directly, you get the unicode version. This eliminates any need to deal with the email.header decode_header() or make_header() functions.

You can also create addresses from parts:

>>> m['cc'] = [Group('pals', [Address('Bob', 'bob', 'example.com'),
...                           Address('Sally', 'sally', 'example.com')]),
...            Address('Bonzo', addr_spec='[email protected]')]
>>> print(m)
To: =?utf-8?q?=C3=89ric?= <[email protected]>
Date: Fri, 25 May 2012 21:44:27 -0400
cc: pals: Bob <[email protected]>, Sally <[email protected]>;, Bonzo <[email protected]>

Decoding to unicode is done automatically:

>>> m2 = message_from_string(str(m))
>>> m2['to']
'Éric <[email protected]>'

When you parse a message, you can use the addresses and groups attributes of the header objects to access the groups and individual addresses:

>>> m2['cc'].addresses
(Address(display_name='Bob', username='bob', domain='example.com'), Address(display_name='Sally', username='sally', domain='example.com'), Address(display_name='Bonzo', username='bonz', domain='laugh.com'))
>>> m2['cc'].groups
(Group(display_name='pals', addresses=(Address(display_name='Bob', username='bob', domain='example.com'), Address(display_name='Sally', username='sally', domain='example.com')), Group(display_name=None, addresses=(Address(display_name='Bonzo', username='bonz', domain='laugh.com'),))

In summary, if you use one of the new policies, header manipulation works the way it ought to: your application works with unicode strings, and the email package transparently encodes and decodes the unicode to and from the RFC standard Content Transfer Encodings.

Other API Changes

New BytesHeaderParser, added to the parser module to complement HeaderParser and complete the Bytes API.

New utility functions:


  • ftplib.FTP now accepts a source_address keyword argument to specify the (host, port) to use as the source address in the bind call when creating the outgoing socket. (Contributed by Giampaolo Rodolà in bpo-8594 .)

  • The FTP_TLS class now provides a new ccc() function to revert control channel back to plaintext. This can be useful to take advantage of firewalls that know how to handle NAT with non-secure FTP without opening fixed ports. (Contributed by Giampaolo Rodolà in bpo-12139 .)

  • Added ftplib.FTP.mlsd() method which provides a parsable directory listing format and deprecates ftplib.FTP.nlst() and ftplib.FTP.dir(). (Contributed by Giampaolo Rodolà in bpo-11072 .)


The functools.lru_cache() decorator now accepts a typed keyword argument (that defaults to False to ensure that it caches values of different types that compare equal in separate cache slots. (Contributed by Raymond Hettinger in bpo-13227 .)


It is now possible to register callbacks invoked by the garbage collector before and after collection using the new callbacks list.


A new compare_digest() function has been added to prevent side channel attacks on digests through timing analysis. (Contributed by Nick Coghlan and Christian Heimes in bpo-15061 .)


http.server.BaseHTTPRequestHandler now buffers the headers and writes them all at once when end_headers() is called. A new method flush_headers() can be used to directly manage when the accumulated headers are sent. (Contributed by Andrew Schaaf in bpo-3709 .)

http.server now produces valid HTML 4.01 strict output. (Contributed by Ezio Melotti in bpo-13295 .)

http.client.HTTPResponse now has a readinto() method, which means it can be used as an io.RawIOBase class. (Contributed by John Kuhn in bpo-13464 .)


html.parser.HTMLParser is now able to parse broken markup without raising errors, therefore the strict argument of the constructor and the HTMLParseError exception are now deprecated. The ability to parse broken markup is the result of a number of bug fixes that are also available on the latest bug fix releases of Python 2.7/3.2. (Contributed by Ezio Melotti in bpo-15114 , and bpo-14538 , bpo-13993 , bpo-13960 , bpo-13358 , bpo-1745761 , bpo-755670 , bpo-13357 , bpo-12629 , bpo-1200313 , bpo-670664 , bpo-13273 , bpo-12888 , bpo-7311 .)

A new html5 dictionary that maps HTML5 named character references to the equivalent Unicode character(s) (e.g. html5['gt;'] == '>') has been added to the html.entities module. The dictionary is now also used by HTMLParser. (Contributed by Ezio Melotti in bpo-11113 and bpo-15156 .)


The IMAP4_SSL constructor now accepts an SSLContext parameter to control parameters of the secure channel.

(Contributed by Sijin Joseph in bpo-8808 .)


A new getclosurevars() function has been added. This function reports the current binding of all names referenced from the function body and where those names were resolved, making it easier to verify correct internal state when testing code that relies on stateful closures.

(Contributed by Meador Inge and Nick Coghlan in bpo-13062 .)

A new getgeneratorlocals() function has been added. This function reports the current binding of local variables in the generator’s stack frame, making it easier to verify correct internal state when testing generators.

(Contributed by Meador Inge in bpo-15153 .)


The open() function has a new 'x' mode that can be used to exclusively create a new file, and raise a FileExistsError if the file already exists. It is based on the C11 ‘x’ mode to fopen().

(Contributed by David Townshend in bpo-12760 .)

The constructor of the TextIOWrapper class has a new write_through optional argument. If write_through is True, calls to write() are guaranteed not to be buffered: any data written on the TextIOWrapper object is immediately handled to its underlying binary buffer.


accumulate() now takes an optional func argument for providing a user-supplied binary function.


The basicConfig() function now supports an optional handlers argument taking an iterable of handlers to be added to the root logger.

A class level attribute append_nul has been added to SysLogHandler to allow control of the appending of the NUL (\000) byte to syslog records, since for some daemons it is required while for others it is passed through to the log.


The math module has a new function, log2(), which returns the base-2 logarithm of x.

(Written by Mark Dickinson in bpo-11888 .)


The read() method is now more compatible with other file-like objects: if the argument is omitted or specified as None, it returns the bytes from the current file position to the end of the mapping. (Contributed by Petri Lehtinen in bpo-12021 .)


The new multiprocessing.connection.wait() function allows polling multiple objects (such as connections, sockets and pipes) with a timeout. (Contributed by Richard Oudkerk in bpo-12328 .)

multiprocessing.Connection objects can now be transferred over multiprocessing connections. (Contributed by Richard Oudkerk in bpo-4892 .)

multiprocessing.Process now accepts a daemon keyword argument to override the default behavior of inheriting the daemon flag from the parent process (bpo-6064 ).

New attribute multiprocessing.Process.sentinel allows a program to wait on multiple Process objects at one time using the appropriate OS primitives (for example, select on posix systems).

New methods multiprocessing.pool.Pool.starmap() and starmap_async() provide itertools.starmap() equivalents to the existing multiprocessing.pool.Pool.map() and map_async() functions. (Contributed by Hynek Schlawack in bpo-12708 .)


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

>>> from nntplib import NNTP
>>> with NNTP('news.gmane.org') as n:
...     n.group('gmane.comp.python.committers')
('211 1755 1 1755 gmane.comp.python.committers', 1755, 1, 1755, 'gmane.comp.python.committers')

(Contributed by Giampaolo Rodolà in bpo-9795 .)


  • The os module has a new pipe2() function that makes it possible to create a pipe with O_CLOEXEC or O_NONBLOCK flags set atomically. This is especially useful to avoid race conditions in multi-threaded programs.

  • The os module has a new sendfile() function which provides an efficient “zero-copy” way for copying data from one file (or socket) descriptor to another. The phrase “zero-copy” refers to the fact that all of the copying of data between the two descriptors is done entirely by the kernel, with no copying of data into userspace buffers. sendfile() can be used to efficiently copy data from a file on disk to a network socket, e.g. for downloading a file.

    (Patch submitted by Ross Lagerwall and Giampaolo Rodolà in bpo-10882 .)

  • To avoid race conditions like symlink attacks and issues with temporary files and directories, it is more reliable (and also faster) to manipulate file descriptors instead of file names. Python 3.3 enhances existing functions and introduces new functions to work on file descriptors (bpo-4761 , bpo-10755 and bpo-14626 ).

  • access() accepts an effective_ids keyword argument to turn on using the effective uid/gid rather than the real uid/gid in the access check. Platform support for this can be checked via the supports_effective_ids set.

  • The os module has two new functions: getpriority() and setpriority(). They can be used to get or set process niceness/priority in a fashion similar to os.nice() but extended to all processes instead of just the current one.

    (Patch submitted by Giampaolo Rodolà in bpo-10784 .)

  • The new os.replace() function allows cross-platform renaming of a file with overwriting the destination. With os.rename(), an existing destination file is overwritten under POSIX, but raises an error under Windows. (Contributed by Antoine Pitrou in bpo-8828 .)

  • The stat family of functions (stat(), fstat(), and lstat()) now support reading a file’s timestamps with nanosecond precision. Symmetrically, utime() can now write file timestamps with nanosecond precision. (Contributed by Larry Hastings in bpo-14127 .)

  • The new os.get_terminal_size() function queries the size of the terminal attached to a file descriptor. See also shutil.get_terminal_size(). (Contributed by Zbigniew Jędrzejewski-Szmek in bpo-13609 .)


Tab-completion is now available not only for command names, but also their arguments. For example, for the break command, function and file names are completed.

(Contributed by Georg Brandl in bpo-14210 )


pickle.Pickler objects now have an optional dispatch_table attribute allowing per-pickler reduction functions to be set.

(Contributed by Richard Oudkerk in bpo-14166 .)


The Tk GUI and the serve() function have been removed from the pydoc module: pydoc -g and serve() have been deprecated in Python 3.2.


str regular expressions now support \u and \U escapes.

(Contributed by Serhiy Storchaka in bpo-3665 .)


  • run() now accepts a blocking parameter which when set to false makes the method execute the scheduled events due to expire soonest (if any) and then return immediately. This is useful in case you