What’s New In Python 3.6

Date:February 15, 2017
Editors:Elvis Pranskevichus <elvis@magic.io>, Yury Selivanov <yury@magic.io>

This article explains the new features in Python 3.6, compared to 3.5. Python 3.6 was released on December 23, 2016.  See the changelog for a full list of changes.

See also

PEP 494 - Python 3.6 Release Schedule

Summary – Release highlights

New syntax features:

  • PEP 498, formatted string literals.
  • PEP 515, underscores in numeric literals.
  • PEP 526, syntax for variable annotations.
  • PEP 525, asynchronous generators.
  • PEP 530: asynchronous comprehensions.

New library modules:

CPython implementation improvements:

Significant improvements in the standard library:

  • The asyncio module has received new features, significant usability and performance improvements, and a fair amount of bug fixes. Starting with Python 3.6 the asyncio module is no longer provisional and its API is considered stable.
  • A new file system path protocol has been implemented to support path-like objects. All standard library functions operating on paths have been updated to work with the new protocol.
  • The datetime module has gained support for Local Time Disambiguation.
  • The typing module received a number of improvements.
  • The tracemalloc module has been significantly reworked and is now used to provide better output for ResourceWarning as well as provide better diagnostics for memory allocation errors. See the PYTHONMALLOC section for more information.

Security improvements:

  • The new secrets module has been added to simplify the generation of cryptographically strong pseudo-random numbers suitable for managing secrets such as account authentication, tokens, and similar.
  • On Linux, os.urandom() now blocks until the system urandom entropy pool is initialized to increase the security. See the PEP 524 for the rationale.
  • The hashlib and ssl modules now support OpenSSL 1.1.0.
  • The default settings and feature set of the ssl module have been improved.
  • The hashlib module received support for the BLAKE2, SHA-3 and SHAKE hash algorithms and the scrypt() key derivation function.

Windows improvements:

  • PEP 528 and PEP 529, Windows filesystem and console encoding changed to UTF-8.
  • The py.exe launcher, when used interactively, no longer prefers Python 2 over Python 3 when the user doesn’t specify a version (via command line arguments or a config file). Handling of shebang lines remains unchanged - “python” refers to Python 2 in that case.
  • python.exe and pythonw.exe have been marked as long-path aware, which means that the 260 character path limit may no longer apply. See removing the MAX_PATH limitation for details.
  • A ._pth file can be added to force isolated mode and fully specify all search paths to avoid registry and environment lookup. See the documentation for more information.
  • A python36.zip file now works as a landmark to infer PYTHONHOME. See the documentation for more information.

New Features

PEP 498: Formatted string literals

PEP 498 introduces a new kind of string literals: f-strings, or formatted string literals.

Formatted string literals are prefixed with 'f' and are similar to the format strings accepted by str.format(). They contain replacement fields surrounded by curly braces. The replacement fields are expressions, which are evaluated at run time, and then formatted using the format() protocol:

>>> name = "Fred"
>>> f"He said his name is {name}."
'He said his name is Fred.'
>>> width = 10
>>> precision = 4
>>> value = decimal.Decimal("12.34567")
>>> f"result: {value:{width}.{precision}}"  # nested fields
'result:      12.35'

See also

PEP 498 – Literal String Interpolation.
PEP written and implemented by Eric V. Smith.

Feature documentation.

PEP 526: Syntax for variable annotations

PEP 484 introduced the standard for type annotations of function parameters, a.k.a. type hints. This PEP adds syntax to Python for annotating the types of variables including class variables and instance variables:

primes: List[int] = []

captain: str  # Note: no initial value!

class Starship:
    stats: Dict[str, int] = {}

Just as for function annotations, the Python interpreter does not attach any particular meaning to variable annotations and only stores them in the __annotations__ attribute of a class or module.

In contrast to variable declarations in statically typed languages, the goal of annotation syntax is to provide an easy way to specify structured type metadata for third party tools and libraries via the abstract syntax tree and the __annotations__ attribute.

See also

PEP 526 – Syntax for variable annotations.
PEP written by Ryan Gonzalez, Philip House, Ivan Levkivskyi, Lisa Roach, and Guido van Rossum. Implemented by Ivan Levkivskyi.

Tools that use or will use the new syntax: mypy, pytype, PyCharm, etc.

PEP 515: Underscores in Numeric Literals

PEP 515 adds the ability to use underscores in numeric literals for improved readability. For example:

>>> 1_000_000_000_000_000
>>> 0x_FF_FF_FF_FF

Single underscores are allowed between digits and after any base specifier. Leading, trailing, or multiple underscores in a row are not allowed.

The string formatting language also now has support for the '_' option to signal the use of an underscore for a thousands separator for floating point presentation types and for integer presentation type 'd'. For integer presentation types 'b', 'o', 'x', and 'X', underscores will be inserted every 4 digits:

>>> '{:_}'.format(1000000)
>>> '{:_x}'.format(0xFFFFFFFF)

See also

PEP 515 – Underscores in Numeric Literals
PEP written by Georg Brandl and Serhiy Storchaka.

PEP 525: Asynchronous Generators

PEP 492 introduced support for native coroutines and async / await syntax to Python 3.5. A notable limitation of the Python 3.5 implementation is that it was not possible to use await and yield in the same function body. In Python 3.6 this restriction has been lifted, making it possible to define asynchronous generators:

async def ticker(delay, to):
    """Yield numbers from 0 to *to* every *delay* seconds."""
    for i in range(to):
        yield i
        await asyncio.sleep(delay)

The new syntax allows for faster and more concise code.

See also

PEP 525 – Asynchronous Generators
PEP written and implemented by Yury Selivanov.

PEP 530: Asynchronous Comprehensions

PEP 530 adds support for using async for in list, set, dict comprehensions and generator expressions:

result = [i async for i in aiter() if i % 2]

Additionally, await expressions are supported in all kinds of comprehensions:

result = [await fun() for fun in funcs if await condition()]

See also

PEP 530 – Asynchronous Comprehensions
PEP written and implemented by Yury Selivanov.

PEP 487: Simpler customization of class creation

It is now possible to customize subclass creation without using a metaclass. The new __init_subclass__ classmethod will be called on the base class whenever a new subclass is created:

class PluginBase:
    subclasses = []

    def __init_subclass__(cls, **kwargs):

class Plugin1(PluginBase):

class Plugin2(PluginBase):

In order to allow zero-argument super() calls to work correctly from __init_subclass__() implementations, custom metaclasses must ensure that the new __classcell__ namespace entry is propagated to type.__new__ (as described in Creating the class object).

See also

PEP 487 – Simpler customization of class creation
PEP written and implemented by Martin Teichmann.

Feature documentation

PEP 487: Descriptor Protocol Enhancements

PEP 487 extends the descriptor protocol to include the new optional __set_name__() method. Whenever a new class is defined, the new method will be called on all descriptors included in the definition, providing them with a reference to the class being defined and the name given to the descriptor within the class namespace. In other words, instances of descriptors can now know the attribute name of the descriptor in the owner class:

class IntField:
    def __get__(self, instance, owner):
        return instance.__dict__[self.name]

    def __set__(self, instance, value):
        if not isinstance(value, int):
            raise ValueError(f'expecting integer in {self.name}')
        instance.__dict__[self.name] = value

    # this is the new initializer:
    def __set_name__(self, owner, name):
        self.name = name

class Model:
    int_field = IntField()

See also

PEP 487 – Simpler customization of class creation
PEP written and implemented by Martin Teichmann.

Feature documentation

PEP 519: Adding a file system path protocol

File system paths have historically been represented as str or bytes objects. This has led to people who write code which operate on file system paths to assume that such objects are only one of those two types (an int representing a file descriptor does not count as that is not a file path). Unfortunately that assumption prevents alternative object representations of file system paths like pathlib from working with pre-existing code, including Python’s standard library.

To fix this situation, a new interface represented by os.PathLike has been defined. By implementing the __fspath__() method, an object signals that it represents a path. An object can then provide a low-level representation of a file system path as a str or bytes object. This means an object is considered path-like if it implements os.PathLike or is a str or bytes object which represents a file system path. Code can use os.fspath(), os.fsdecode(), or os.fsencode() to explicitly get a str and/or bytes representation of a path-like object.

The built-in open() function has been updated to accept os.PathLike objects, as have all relevant functions in the os and os.path modules, and most other functions and classes in the standard library. The os.DirEntry class and relevant classes in pathlib have also been updated to implement os.PathLike.

The hope is that updating the fundamental functions for operating on file system paths will lead to third-party code to implicitly support all path-like objects without any code changes, or at least very minimal ones (e.g. calling os.fspath() at the beginning of code before operating on a path-like object).

Here are some examples of how the new interface allows for pathlib.Path to be used more easily and transparently with pre-existing code:

>>> import pathlib
>>> with open(pathlib.Path("README")) as f:
...     contents = f.read()
>>> import os.path
>>> os.path.splitext(pathlib.Path("some_file.txt"))
('some_file', '.txt')
>>> os.path.join("/a/b", pathlib.Path("c"))
>>> import os
>>> os.fspath(pathlib.Path("some_file.txt"))

(Implemented by Brett Cannon, Ethan Furman, Dusty Phillips, and Jelle Zijlstra.)

See also

PEP 519 – Adding a file system path protocol
PEP written by Brett Cannon and Koos Zevenhoven.

PEP 495: Local Time Disambiguation

In most world locations, there have been and will be times when local clocks are moved back. In those times, intervals are introduced in which local clocks show the same time twice in the same day. In these situations, the information displayed on a local clock (or stored in a Python datetime instance) is insufficient to identify a particular moment in time.

PEP 495 adds the new fold attribute to instances of datetime.datetime and datetime.time classes to differentiate between two moments in time for which local times are the same:

>>> u0 = datetime(2016, 11, 6, 4, tzinfo=timezone.utc)
>>> for i in range(4):
...     u = u0 + i*HOUR
...     t = u.astimezone(Eastern)
...     print(u.time(), 'UTC =', t.time(), t.tzname(), t.fold)
04:00:00 UTC = 00:00:00 EDT 0
05:00:00 UTC = 01:00:00 EDT 0
06:00:00 UTC = 01:00:00 EST 1
07:00:00 UTC = 02:00:00 EST 0

The values of the fold attribute have the value 0 for all instances except those that represent the second (chronologically) moment in time in an ambiguous case.

See also

PEP 495 – Local Time Disambiguation
PEP written by Alexander Belopolsky and Tim Peters, implementation by Alexander Belopolsky.

PEP 529: Change Windows filesystem encoding to UTF-8

Representing filesystem paths is best performed with str (Unicode) rather than bytes. However, there are some situations where using bytes is sufficient and correct.

Prior to Python 3.6, data loss could result when using bytes paths on Windows. With this change, using bytes to represent paths is now supported on Windows, provided those bytes are encoded with the encoding returned by sys.getfilesystemencoding(), which now defaults to 'utf-8'.

Applications that do not use str to represent paths should use os.fsencode() and os.fsdecode() to ensure their bytes are correctly encoded. To revert to the previous behaviour, set PYTHONLEGACYWINDOWSFSENCODING or call sys._enablelegacywindowsfsencoding().

See PEP 529 for more information and discussion of code modifications that may be required.

PEP 528: Change Windows console encoding to UTF-8

The default console on Windows will now accept all Unicode characters and provide correctly read str objects to Python code. sys.stdin, sys.stdout and sys.stderr now default to utf-8 encoding.

This change only applies when using an interactive console, and not when redirecting files or pipes. To revert to the previous behaviour for interactive console use, set PYTHONLEGACYWINDOWSIOENCODING.

See also

PEP 528 – Change Windows console encoding to UTF-8
PEP written and implemented by Steve Dower.

PEP 520: Preserving Class Attribute Definition Order

Attributes in a class definition body have a natural ordering: the same order in which the names appear in the source. This order is now preserved in the new class’s __dict__ attribute.

Also, the effective default class execution namespace (returned from type.__prepare__()) is now an insertion-order-preserving mapping.

See also

PEP 520 – Preserving Class Attribute Definition Order
PEP written and implemented by Eric Snow.

PEP 468: Preserving Keyword Argument Order

**kwargs in a function signature is now guaranteed to be an insertion-order-preserving mapping.

See also

PEP 468 – Preserving Keyword Argument Order
PEP written and implemented by Eric Snow.

New dict implementation

The dict type now uses a “compact” representation based on a proposal by Raymond Hettinger which was first implemented by PyPy. The memory usage of the new dict() is between 20% and 25% smaller compared to Python 3.5.

The order-preserving aspect of this new implementation is considered an implementation detail and should not be relied upon (this may change in the future, but it is desired to have this new dict implementation in the language for a few releases before changing the language spec to mandate order-preserving semantics for all current and future Python implementations; this also helps preserve backwards-compatibility with older versions of the language where random iteration order is still in effect, e.g. Python 3.5).

(Contributed by INADA Naoki in bpo-27350. Idea originally suggested by Raymond Hettinger.)

PEP 523: Adding a frame evaluation API to CPython

While Python provides extensive support to customize how code executes, one place it has not done so is in the evaluation of frame objects. If you wanted some way to intercept frame evaluation in Python there really wasn’t any way without directly manipulating function pointers for defined functions.

PEP 523 changes this by providing an API to make frame evaluation pluggable at the C level. This will allow for tools such as debuggers and JITs to intercept frame evaluation before the execution of Python code begins. This enables the use of alternative evaluation implementations for Python code, tracking frame evaluation, etc.

This API is not part of the limited C API and is marked as private to signal that usage of this API is expected to be limited and only applicable to very select, low-level use-cases. Semantics of the API will change with Python as necessary.

See also

PEP 523 – Adding a frame evaluation API to CPython
PEP written by Brett Cannon and Dino Viehland.

PYTHONMALLOC environment variable

The new PYTHONMALLOC environment variable allows setting the Python memory allocators and installing debug hooks.

It is now possible to install debug hooks on Python memory allocators on Python compiled in release mode using PYTHONMALLOC=debug. Effects of debug hooks:

  • Newly allocated memory is filled with the byte 0xCB
  • Freed memory is filled with the byte 0xDB
  • Detect violations of the Python memory allocator API. For example, PyObject_Free() called on a memory block allocated by PyMem_Malloc().
  • Detect writes before the start of a buffer (buffer underflows)
  • Detect writes after the end of a buffer (buffer overflows)
  • Check that the GIL is held when allocator functions of PYMEM_DOMAIN_OBJ (ex: PyObject_Malloc()) and PYMEM_DOMAIN_MEM (ex: PyMem_Malloc()) domains are called.

Checking if the GIL is held is also a new feature of Python 3.6.

See the PyMem_SetupDebugHooks() function for debug hooks on Python memory allocators.

It is now also possible to force the usage of the malloc() allocator of the C library for all Python memory allocations using PYTHONMALLOC=malloc. This is helpful when using external memory debuggers like Valgrind on a Python compiled in release mode.

On error, the debug hooks on Python memory allocators now use the tracemalloc module to get the traceback where a memory block was allocated.

Example of fatal error on buffer overflow using python3.6 -X tracemalloc=5 (store 5 frames in traces):

Debug memory block at address p=0x7fbcd41666f8: API 'o'
    4 bytes originally requested
    The 7 pad bytes at p-7 are FORBIDDENBYTE, as expected.
    The 8 pad bytes at tail=0x7fbcd41666fc are not all FORBIDDENBYTE (0xfb):
        at tail+0: 0x02 *** OUCH
        at tail+1: 0xfb
        at tail+2: 0xfb
        at tail+3: 0xfb
        at tail+4: 0xfb
        at tail+5: 0xfb
        at tail+6: 0xfb
        at tail+7: 0xfb
    The block was made by call #1233329 to debug malloc/realloc.
    Data at p: 1a 2b 30 00

Memory block allocated at (most recent call first):
  File "test/test_bytes.py", line 323
  File "unittest/case.py", line 600
  File "unittest/case.py", line 648
  File "unittest/suite.py", line 122
  File "unittest/suite.py", line 84

Fatal Python error: bad trailing pad byte

Current thread 0x00007fbcdbd32700 (most recent call first):
  File "test/test_bytes.py", line 323 in test_hex
  File "unittest/case.py", line 600 in run
  File "unittest/case.py", line 648 in __call__
  File "unittest/suite.py", line 122 in run
  File "unittest/suite.py", line 84 in __call__
  File "unittest/suite.py", line 122 in run
  File "unittest/suite.py", line 84 in __call__

(Contributed by Victor Stinner in bpo-26516 and bpo-26564.)

DTrace and SystemTap probing support

Python can now be built --with-dtrace which enables static markers for the following events in the interpreter:

  • function call/return
  • garbage collection started/finished
  • line of code executed.

This can be used to instrument running interpreters in production, without the need to recompile specific debug builds or providing application-specific profiling/debugging code.

More details in Instrumenting CPython with DTrace and SystemTap.

The current implementation is tested on Linux and macOS. Additional markers may be added in the future.

(Contributed by Łukasz Langa in bpo-21590, based on patches by Jesús Cea Avión, David Malcolm, and Nikhil Benesch.)

Other Language Changes

Some smaller changes made to the core Python language are:

  • A global or nonlocal statement must now textually appear before the first use of the affected name in the same scope. Previously this was a SyntaxWarning.
  • It is now possible to set a special method to None to indicate that the corresponding operation is not available. For example, if a class sets __iter__() to None, the class is not iterable. (Contributed by Andrew Barnert and Ivan Levkivskyi in bpo-25958.)
  • Long sequences of repeated traceback lines are now abbreviated as "[Previous line repeated {count} more times]" (see traceback for an example). (Contributed by Emanuel Barry in bpo-26823.)
  • Import now raises the new exception ModuleNotFoundError (subclass of ImportError) when it cannot find a module. Code that currently checks for ImportError (in try-except) will still work. (Contributed by Eric Snow in bpo-15767.)
  • Class methods relying on zero-argument super() will now work correctly when called from metaclass methods during class creation. (Contributed by Martin Teichmann in bpo-23722.)

New Modules


The main purpose of the new secrets module is to provide an obvious way to reliably generate cryptographically strong pseudo-random values suitable for managing secrets, such as account authentication, tokens, and similar.


Note that the pseudo-random generators in the random module should NOT be used for security purposes. Use secrets on Python 3.6+ and os.urandom() on Python 3.5 and earlier.

See also

PEP 506 – Adding A Secrets Module To The Standard Library
PEP written and implemented by Steven D’Aprano.

Improved Modules


Exhausted iterators of array.array will now stay exhausted even if the iterated array is extended. This is consistent with the behavior of other mutable sequences.

Contributed by Serhiy Storchaka in bpo-26492.


The new ast.Constant AST node has been added. It can be used by external AST optimizers for the purposes of constant folding.

Contributed by Victor Stinner in bpo-26146.


Starting with Python 3.6 the asyncio module is no longer provisional and its API is considered stable.

Notable changes in the asyncio module since Python 3.5.0 (all backported to 3.5.x due to the provisional status):

  • The get_event_loop() function has been changed to always return the currently running loop when called from couroutines and callbacks. (Contributed by Yury Selivanov in bpo-28613.)
  • The ensure_future() function and all functions that use it, such as loop.run_until_complete(), now accept all kinds of awaitable objects. (Contributed by Yury Selivanov.)
  • New run_coroutine_threadsafe() function to submit coroutines to event loops from other threads. (Contributed by Vincent Michel.)
  • New Transport.is_closing() method to check if the transport is closing or closed. (Contributed by Yury Selivanov.)
  • The loop.create_server() method can now accept a list of hosts. (Contributed by Yann Sionneau.)
  • New loop.create_future() method to create Future objects. This allows alternative event loop implementations, such as uvloop, to provide a faster asyncio.Future implementation. (Contributed by Yury Selivanov in bpo-27041.)
  • New loop.get_exception_handler() method to get the current exception handler. (Contributed by Yury Selivanov in bpo-27040.)
  • New StreamReader.readuntil() method to read data from the stream until a separator bytes sequence appears. (Contributed by Mark Korenberg.)
  • The performance of StreamReader.readexactly() has been improved. (Contributed by Mark Korenberg in bpo-28370.)
  • The loop.getaddrinfo() method is optimized to avoid calling the system getaddrinfo function if the address is already resolved. (Contributed by A. Jesse Jiryu Davis.)
  • The loop.stop() method has been changed to stop the loop immediately after the current iteration. Any new callbacks scheduled as a result of the last iteration will be discarded. (Contributed by Guido van Rossum in bpo-25593.)
  • Future.set_exception will now raise TypeError when passed an instance of the StopIteration exception. (Contributed by Chris Angelico in bpo-26221.)
  • New loop.connect_accepted_socket() method to be used by servers that accept connections outside of asyncio, but that use asyncio to handle them. (Contributed by Jim Fulton in bpo-27392.)
  • TCP_NODELAY flag is now set for all TCP transports by default. (Contributed by Yury Selivanov in bpo-27456.)
  • New loop.shutdown_asyncgens() to properly close pending asynchronous generators before closing the loop. (Contributed by Yury Selivanov in bpo-28003.)
  • Future and Task classes now have an optimized C implementation which makes asyncio code up to 30% faster. (Contributed by Yury Selivanov and INADA Naoki in bpo-26081 and bpo-28544.)


The b2a_base64() function now accepts an optional newline keyword argument to control whether the newline character is appended to the return value. (Contributed by Victor Stinner in bpo-25357.)


The new cmath.tau (τ) constant has been added. (Contributed by Lisa Roach in bpo-12345, see PEP 628 for details.)

New constants: cmath.inf and cmath.nan to match math.inf and math.nan, and also cmath.infj and cmath.nanj to match the format used by complex repr. (Contributed by Mark Dickinson in bpo-23229.)


The new Collection abstract base class has been added to represent sized iterable container classes. (Contributed by Ivan Levkivskyi, docs by Neil Girdhar in bpo-27598.)

The new Reversible abstract base class represents iterable classes that also provide the __reversed__() method. (Contributed by Ivan Levkivskyi in bpo-25987.)

The new AsyncGenerator abstract base class represents asynchronous generators. (Contributed by Yury Selivanov in bpo-28720.)

The namedtuple() function now accepts an optional keyword argument module, which, when specified, is used for the __module__ attribute of the returned named tuple class. (Contributed by Raymond Hettinger in bpo-17941.)

The verbose and rename arguments for namedtuple() are now keyword-only. (Contributed by Raymond Hettinger in bpo-25628.)

Recursive collections.deque instances can now be pickled. (Contributed by Serhiy Storchaka in bpo-26482.)


The ThreadPoolExecutor class constructor now accepts an optional thread_name_prefix argument to make it possible to customize the names of the threads created by the pool. (Contributed by Gregory P. Smith in bpo-27664.)


The contextlib.AbstractContextManager class has been added to provide an abstract base class for context managers. It provides a sensible default implementation for __enter__() which returns self and leaves __exit__() an abstract method. A matching class has been added to the typing module as typing.ContextManager. (Contributed by Brett Cannon in bpo-25609.)


The datetime and time classes have the new fold attribute used to disambiguate local time when necessary. Many functions in the datetime have been updated to support local time disambiguation. See Local Time Disambiguation section for more information. (Contributed by Alexander Belopolsky in bpo-24773.)

The datetime.strftime() and date.strftime() methods now support ISO 8601 date directives %G, %u and %V. (Contributed by Ashley Anderson in bpo-12006.)

The datetime.isoformat() function now accepts an optional timespec argument that specifies the number of additional components of the time value to include. (Contributed by Alessandro Cucci and Alexander Belopolsky in bpo-19475.)

The datetime.combine() now accepts an optional tzinfo argument. (Contributed by Alexander Belopolsky in bpo-27661.)


New Decimal.as_integer_ratio() method that returns a pair (n, d) of integers that represent the given Decimal instance as a fraction, in lowest terms and with a positive denominator:

>>> Decimal('-3.14').as_integer_ratio()
(-157, 50)

(Contributed by Stefan Krah amd Mark Dickinson in bpo-25928.)


The default_format attribute has been removed from distutils.command.sdist.sdist and the formats attribute defaults to ['gztar']. Although not anticipated, any code relying on the presence of default_format may need to be adapted. See bpo-27819 for more details.


The new email API, enabled via the policy keyword to various constructors, is no longer provisional. The email documentation has been reorganized and rewritten to focus on the new API, while retaining the old documentation for the legacy API. (Contributed by R. David Murray in bpo-24277.)

The email.mime classes now all accept an optional policy keyword. (Contributed by Berker Peksag in bpo-27331.)

The DecodedGenerator now supports the policy keyword.

There is a new policy attribute, message_factory, that controls what class is used by default when the parser creates new message objects. For the email.policy.compat32 policy this is Message, for the new policies it is EmailMessage. (Contributed by R. David Murray in bpo-20476.)


On Windows, added the 'oem' encoding to use CP_OEMCP, and the 'ansi' alias for the existing 'mbcs' encoding, which uses the CP_ACP code page. (Contributed by Steve Dower in bpo-27959.)


Two new enumeration base classes have been added to the enum module: Flag and IntFlags. Both are used to define constants that can be combined using the bitwise operators. (Contributed by Ethan Furman in bpo-23591.)

Many standard library modules have been updated to use the IntFlags class for their constants.

The new enum.auto value can be used to assign values to enum members automatically:

>>> from enum import Enum, auto
>>> class Color(Enum):
...     red = auto()
...     blue = auto()
...     green = auto()
>>> list(Color)
[<Color.red: 1>, <Color.blue: 2>, <Color.green: 3>]


On Windows, the faulthandler module now installs a handler for Windows exceptions: see faulthandler.enable(). (Contributed by Victor Stinner in bpo-23848.)


hook_encoded() now supports the errors argument. (Contributed by Joseph Hackman in bpo-25788.)


hashlib supports OpenSSL 1.1.0. The minimum recommend version is 1.0.2. (Contributed by Christian Heimes in bpo-26470.)

BLAKE2 hash functions were added to the module. blake2b() and blake2s() are always available and support the full feature set of BLAKE2. (Contributed by Christian Heimes in bpo-26798 based on code by Dmitry Chestnykh and Samuel Neves. Documentation written by Dmitry Chestnykh.)

The SHA-3 hash functions sha3_224(), sha3_256(), sha3_384(), sha3_512(), and SHAKE hash functions shake_128() and shake_256() were added. (Contributed by Christian Heimes in bpo-16113. Keccak Code Package by Guido Bertoni, Joan Daemen, Michaël Peeters, Gilles Van Assche, and Ronny Van Keer.)

The password-based key derivation function scrypt() is now available with OpenSSL 1.1.0 and newer. (Contributed by Christian Heimes in bpo-27928.)


HTTPConnection.request() and endheaders() both now support chunked encoding request bodies. (Contributed by Demian Brecht and Rolf Krahl in bpo-12319.)

idlelib and IDLE

The idlelib package is being modernized and refactored to make IDLE look and work better and to make the code easier to understand, test, and improve. Part of making IDLE look better, especially on Linux and Mac, is using ttk widgets, mostly in the dialogs. As a result, IDLE no longer runs with tcl/tk 8.4. It now requires tcl/tk 8.5 or 8.6. We recommend running the latest release of either.

‘Modernizing’ includes renaming and consolidation of idlelib modules. The renaming of files with partial uppercase names is similar to the renaming of, for instance, Tkinter and TkFont to tkinter and tkinter.font in 3.0. As a result, imports of idlelib files that worked in 3.5 will usually not work in 3.6. At least a module name change will be needed (see idlelib/README.txt), sometimes more. (Name changes contributed by Al Swiegart and Terry Reedy in bpo-24225. Most idlelib patches since have been and will be part of the process.)

In compensation, the eventual result with be that some idlelib classes will be easier to use, with better APIs and docstrings explaining them. Additional useful information will be added to idlelib when available.


Import now raises the new exception ModuleNotFoundError (subclass of ImportError) when it cannot find a module. Code that current checks for ImportError (in try-except) will still work. (Contributed by Eric Snow in bpo-15767.)

importlib.util.LazyLoader now calls create_module() on the wrapped loader, removing the restriction that importlib.machinery.BuiltinImporter and importlib.machinery.ExtensionFileLoader couldn’t be used with importlib.util.LazyLoader.

importlib.util.cache_from_source(), importlib.util.source_from_cache(), and importlib.util.spec_from_file_location() now accept a path-like object.


The inspect.signature() function now reports the implicit .0 parameters generated by the compiler for comprehension and generator expression scopes as if they were positional-only parameters called implicit0. (Contributed by Jelle Zijlstra in bpo-19611.)

To reduce code churn when upgrading from Python 2.7 and the legacy inspect.getargspec() API, the previously documented deprecation of inspect.getfullargspec() has been reversed. While this function is convenient for single/source Python 2/3 code bases, the richer inspect.signature() interface remains the recommended approach for new code. (Contributed by Nick Coghlan in bpo-27172)


json.load() and json.loads() now support binary input. Encoded JSON should be represented using either UTF-8, UTF-16, or UTF-32. (Contributed by Serhiy Storchaka in bpo-17909.)


The new WatchedFileHandler.reopenIfNeeded() method has been added to add the ability to check if the log file needs to be reopened. (Contributed by Marian Horban in bpo-24884.)


The tau (τ) constant has been added to the math and cmath modules. (Contributed by Lisa Roach in bpo-12345, see PEP 628 for details.)


Proxy Objects returned by multiprocessing.Manager() can now be nested. (Contributed by Davin Potts in bpo-6766.)


See the summary of PEP 519 for details on how the os and os.path modules now support path-like objects.

scandir() now supports bytes paths on Windows.

A new close() method allows explicitly closing a scandir() iterator. The scandir() iterator now supports the context manager protocol. If a scandir() iterator is neither exhausted nor explicitly closed a ResourceWarning will be emitted in its destructor. (Contributed by Serhiy Storchaka in bpo-25994.)

On Linux, os.urandom() now blocks until the system urandom entropy pool is initialized to increase the security. See the PEP 524 for the rationale.

The Linux getrandom() syscall (get random bytes) is now exposed as the new os.getrandom() function. (Contributed by Victor Stinner, part of the PEP 524)


pathlib now supports path-like objects. (Contributed by Brett Cannon in bpo-27186.)

See the summary of PEP 519 for details.


The Pdb class constructor has a new optional readrc argument to control whether .pdbrc files should be read.


Objects that need __new__ called with keyword arguments can now be pickled using pickle protocols older than protocol version 4. Protocol version 4 already supports this case. (Contributed by Serhiy Storchaka in bpo-24164.)


pickletools.dis() now outputs the implicit memo index for the MEMOIZE opcode. (Contributed by Serhiy Storchaka in bpo-25382.)


The pydoc module has learned to respect the MANPAGER environment variable. (Contributed by Matthias Klose in bpo-8637.)

help() and pydoc can now list named tuple fields in the order they were defined rather than alphabetically. (Contributed by Raymond Hettinger in bpo-24879.)


The new choices() function returns a list of elements of specified size from the given population with optional weights. (Contributed by Raymond Hettinger in bpo-18844.)


Added support of modifier spans in regular expressions. Examples: '(?i:p)ython' matches 'python' and 'Python', but not 'PYTHON'; '(?i)g(?-i:v)r' matches 'GvR' and 'gvr', but not 'GVR'. (Contributed by Serhiy Storchaka in bpo-433028.)

Match object groups can be accessed by __getitem__, which is equivalent to group(). So mo['name'] is now equivalent to mo.group('name'). (Contributed by Eric Smith in bpo-24454.)

Match objects now support index-like objects as group indices. (Contributed by Jeroen Demeyer and Xiang Zhang in bpo-27177.)


Added set_auto_history() to enable or disable automatic addition of input to the history list. (Contributed by Tyler Crompton in bpo-26870.)


Private and special attribute names now are omitted unless the prefix starts with underscores. A space or a colon is added after some completed keywords. (Contributed by Serhiy Storchaka in bpo-25011 and bpo-25209.)


The shlex has much improved shell compatibility through the new punctuation_chars argument to control which characters are treated as punctuation. (Contributed by Vinay Sajip in bpo-1521950.)


When specifying paths to add to sys.path in a .pth file, you may now specify file paths on top of directories (e.g. zip files). (Contributed by Wolfgang Langner in bpo-26587).


sqlite3.Cursor.lastrowid now supports the REPLACE statement. (Contributed by Alex LordThorsen in bpo-16864.)


The ioctl() function now supports the SIO_LOOPBACK_FAST_PATH control code. (Contributed by Daniel Stokes in bpo-26536.)

The getsockopt() constants SO_DOMAIN, SO_PROTOCOL, SO_PEERSEC, and SO_PASSSEC are now supported. (Contributed by Christian Heimes in bpo-26907.)

The setsockopt() now supports the setsockopt(level, optname, None, optlen: int) form. (Contributed by Christian Heimes in bpo-27744.)

The socket module now supports the address family AF_ALG to interface with Linux Kernel crypto API. ALG_*, SOL_ALG and sendmsg_afalg() were added. (Contributed by Christian Heimes in bpo-27744 with support from Victor Stinner.)

New Linux constants TCP_USER_TIMEOUT and TCP_CONGESTION were added. (Contributed by Omar Sandoval, issue:26273).


Servers based on the socketserver module, including those defined in http.server, xmlrpc.server and wsgiref.simple_server, now support the context manager protocol. (Contributed by Aviv Palivoda in bpo-26404.)

The wfile attribute of StreamRequestHandler classes now implements the io.BufferedIOBase writable interface. In particular, calling write() is now guaranteed to send the data in full. (Contributed by Martin Panter in bpo-26721.)


ssl supports OpenSSL 1.1.0. The minimum recommend version is 1.0.2. (Contributed by Christian Heimes in bpo-26470.)

3DES has been removed from the default cipher suites and ChaCha20 Poly1305 cipher suites have been added. (Contributed by Christian Heimes in bpo-27850 and bpo-27766.)

SSLContext has better default configuration for options and ciphers. (Contributed by Christian Heimes in bpo-28043.)

SSL session can be copied from one client-side connection to another with the new SSLSession class. TLS session resumption can speed up the initial handshake, reduce latency and improve performance (Contributed by Christian Heimes in bpo-19500 based on a draft by Alex Warhawk.)

The new get_ciphers() method can be used to get a list of enabled ciphers in order of cipher priority.

All constants and flags have been converted to IntEnum and IntFlags. (Contributed by Christian Heimes in bpo-28025.)

Server and client-side specific TLS protocols for SSLContext were added. (Contributed by Christian Heimes in bpo-28085.)


A new harmonic_mean() function has been added. (Contributed by Steven D’Aprano in bpo-27181.)


struct now supports IEEE 754 half-precision floats via the 'e' format specifier. (Contributed by Eli Stevens, Mark Dickinson in bpo-11734.)


subprocess.Popen destructor now emits a ResourceWarning warning if the child process is still running. Use the context manager protocol (with proc: ...) or explicitly call the wait() method to read the exit status of the child process. (Contributed by Victor Stinner in bpo-26741.)

The subprocess.Popen constructor and all functions that pass arguments through to it now accept encoding and errors arguments. Specifying either of these will enable text mode for the stdin, stdout and stderr streams. (Contributed by Steve Dower in bpo-6135.)


The new getfilesystemencodeerrors() function returns the name of the error mode used to convert between Unicode filenames and bytes filenames. (Contributed by Steve Dower in bpo-27781.)

On Windows the return value of the getwindowsversion() function now includes the platform_version field which contains the accurate major version, minor version and build number of the current operating system, rather than the version that is being emulated for the process (Contributed by Steve Dower in bpo-27932.)


Telnet is now a context manager (contributed by Stéphane Wirtel in bpo-25485).


The struct_time attributes tm_gmtoff and tm_zone are now available on all platforms.


The new Timer.autorange() convenience method has been added to call Timer.timeit() repeatedly so that the total run time is greater or equal to 200 milliseconds. (Contributed by Steven D’Aprano in bpo-6422.)

timeit now warns when there is substantial (4x) variance between best and worst times. (Contributed by Serhiy Storchaka in bpo-23552.)


Added methods trace_add(), trace_remove() and trace_info() in the tkinter.Variable class. They replace old methods trace_variable(), trace(), trace_vdelete() and trace_vinfo() that use obsolete Tcl commands and might not work in future versions of Tcl. (Contributed by Serhiy Storchaka in bpo-22115).


Both the traceback module and the interpreter’s builtin exception display now abbreviate long sequences of repeated lines in tracebacks as shown in the following example:

>>> def f(): f()
>>> f()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 1, in f
  File "<stdin>", line 1, in f
  File "<stdin>", line 1, in f
  [Previous line repeated 995 more times]
RecursionError: maximum recursion depth exceeded

(Contributed by Emanuel Barry in bpo-26823.)


The tracemalloc module now supports tracing memory allocations in multiple different address spaces.

The new DomainFilter filter class has been added to filter block traces by their address space (domain).

(Contributed by Victor Stinner in bpo-26588.)


Since the typing module is provisional, all changes introduced in Python 3.6 have also been backported to Python 3.5.x.

The typing module has a much improved support for generic type aliases. For example Dict[str, Tuple[S, T]] is now a valid type annotation. (Contributed by Guido van Rossum in Github #195.)

The typing.ContextManager class has been added for representing contextlib.AbstractContextManager. (Contributed by Brett Cannon in bpo-25609.)

The typing.Collection class has been added for representing collections.abc.Collection. (Contributed by Ivan Levkivskyi in bpo-27598.)

The typing.ClassVar type construct has been added to mark class variables. As introduced in PEP 526, a variable annotation wrapped in ClassVar indicates that a given attribute is intended to be used as a class variable and should not be set on instances of that class. (Contributed by Ivan Levkivskyi in Github #280.)

A new TYPE_CHECKING constant that is assumed to be True by the static type chekers, but is False at runtime. (Contributed by Guido van Rossum in Github #230.)

A new NewType() helper function has been added to create lightweight distinct types for annotations:

from typing import NewType

UserId = NewType('UserId', int)
some_id = UserId(524313)

The static type checker will treat the new type as if it were a subclass of the original type. (Contributed by Ivan Levkivskyi in Github #189.)


The unicodedata module now uses data from Unicode 9.0.0. (Contributed by Benjamin Peterson.)


The Mock class has the following improvements:


If a HTTP request has a file or iterable body (other than a bytes object) but no Content-Length header, rather than throwing an error, AbstractHTTPHandler now falls back to use chunked transfer encoding. (Contributed by Demian Brecht and Rolf Krahl in bpo-12319.)


RobotFileParser now supports the Crawl-delay and Request-rate extensions. (Contributed by Nikolay Bogoychev in bpo-16099.)


venv accepts a new parameter --prompt. This parameter provides an alternative prefix for the virtual environment. (Proposed by Łukasz Balcerzak and ported to 3.6 by Stéphane Wirtel in bpo-22829.)


A new optional source parameter has been added to the warnings.warn_explicit() function: the destroyed object which emitted a ResourceWarning. A source attribute has also been added to warnings.WarningMessage (contributed by Victor Stinner in bpo-26568 and bpo-26567).

When a ResourceWarning warning is logged, the tracemalloc module is now used to try to retrieve the traceback where the destroyed object was allocated.

Example with the script example.py:

import warnings

def func():
    return open(__file__)

f = func()
f = None

Output of the command python3.6 -Wd -X tracemalloc=5 example.py:

example.py:7: ResourceWarning: unclosed file <_io.TextIOWrapper name='example.py' mode='r' encoding='UTF-8'>
  f = None
Object allocated at (most recent call first):
  File "example.py", lineno 4
    return open(__file__)
  File "example.py", lineno 6
    f = func()

The “Object allocated at” traceback is new and is only displayed if tracemalloc is tracing Python memory allocations and if the warnings module was already imported.


Added the 64-bit integer type REG_QWORD. (Contributed by Clement Rouault in bpo-23026.)


Allowed keyword arguments to be passed to Beep, MessageBeep, and PlaySound (bpo-27982).


The xmlrpc.client module now supports unmarshalling additional data types used by the Apache XML-RPC implementation for numerics and None. (Contributed by Serhiy Storchaka in bpo-26885.)


A new ZipInfo.from_file() class method allows making a ZipInfo instance from a filesystem file. A new ZipInfo.is_dir() method can be used to check if the ZipInfo instance represents a directory. (Contributed by Thomas Kluyver in bpo-26039.)

The ZipFile.open() method can now be used to write data into a ZIP file, as well as for extracting data. (Contributed by Thomas Kluyver in bpo-26039.)


The compress() and decompress() functions now accept keyword arguments. (Contributed by Aviv Palivoda in bpo-26243 and Xiang Zhang in bpo-16764 respectively.)


  • The Python interpreter now uses a 16-bit wordcode instead of bytecode which made a number of opcode optimizations possible. (Contributed by Demur Rumed with input and reviews from Serhiy Storchaka and Victor Stinner in bpo-26647 and bpo-28050.)
  • The asyncio.Future class now has an optimized C implementation. (Contributed by Yury Selivanov and INADA Naoki in bpo-26081.)
  • The asyncio.Task class now has an optimized C implementation. (Contributed by Yury Selivanov in bpo-28544.)
  • Various implementation improvements in the typing module (such as caching of generic types) allow up to 30 times performance improvements and reduced memory footprint.
  • The ASCII decoder is now up to 60 times as fast for error handlers surrogateescape, ignore and replace (Contributed by Victor Stinner in bpo-24870).
  • The ASCII and the Latin1 encoders are now up to 3 times as fast for the error handler surrogateescape (Contributed by Victor Stinner in bpo-25227).
  • The UTF-8 encoder is now up to 75 times as fast for error handlers ignore, replace, surrogateescape, surrogatepass (Contributed by Victor Stinner in bpo-25267).
  • The UTF-8 decoder is now up to 15 times as fast for error handlers ignore, replace and surrogateescape (Contributed by Victor Stinner in bpo-25301).
  • bytes % args is now up to 2 times faster. (Contributed by Victor Stinner in bpo-25349).
  • bytearray % args is now between 2.5 and 5 times faster. (Contributed by Victor Stinner in bpo-25399).
  • Optimize bytes.fromhex() and bytearray.fromhex(): they are now between 2x and 3.5x faster. (Contributed by Victor Stinner in bpo-25401).
  • Optimize bytes.replace(b'', b'.') and bytearray.replace(b'', b'.'): up to 80% faster. (Contributed by Josh Snider in bpo-26574).
  • Allocator functions of the PyMem_Malloc() domain (PYMEM_DOMAIN_MEM) now use the pymalloc memory allocator instead of malloc() function of the C library. The pymalloc allocator is optimized for objects smaller or equal to 512 bytes with a short lifetime, and use malloc() for larger memory blocks. (Contributed by Victor Stinner in bpo-26249).
  • pickle.load() and pickle.loads() are now up to 10% faster when deserializing many small objects (Contributed by Victor Stinner in bpo-27056).
  • Passing keyword arguments to a function has an overhead in comparison with passing positional arguments. Now in extension functions implemented with using Argument Clinic this overhead is significantly decreased. (Contributed by Serhiy Storchaka in bpo-27574).
  • Optimized glob() and iglob() functions in the glob module; they are now about 3–6 times faster. (Contributed by Serhiy Storchaka in bpo-25596).
  • Optimized globbing in pathlib by using os.scandir(); it is now about 1.5–4 times faster. (Contributed by Serhiy Storchaka in bpo-26032).
  • xml.etree.ElementTree parsing, iteration and deepcopy performance has been significantly improved. (Contributed by Serhiy Storchaka in bpo-25638, bpo-25873, and bpo-25869.)
  • Creation of fractions.Fraction instances from floats and decimals is now 2 to 3 times faster. (Contributed by Serhiy Storchaka in bpo-25971.)

Build and C API Changes

  • Python now requires some C99 support in the toolchain to build. Most notably, Python now uses standard integer types and macros in place of custom macros like PY_LONG_LONG. For more information, see PEP 7 and bpo-17884.
  • Cross-compiling CPython with the Android NDK and the Android API level set to 21 (Android 5.0 Lollilop) or greater runs successfully. While Android is not yet a supported platform, the Python test suite runs on the Android emulator with only about 16 tests failures. See the Android meta-issue bpo-26865.
  • The --enable-optimizations configure flag has been added. Turning it on will activate expensive optimizations like PGO. (Original patch by Alecsandru Patrascu of Intel in bpo-26359.)
  • The GIL must now be held when allocator functions of PYMEM_DOMAIN_OBJ (ex: PyObject_Malloc()) and PYMEM_DOMAIN_MEM (ex: PyMem_Malloc()) domains are called.
  • New Py_FinalizeEx() API which indicates if flushing buffered data failed. (Contributed by Martin Panter in bpo-5319.)
  • PyArg_ParseTupleAndKeywords() now supports positional-only parameters. Positional-only parameters are defined by empty names. (Contributed by Serhiy Storchaka in bpo-26282).
  • PyTraceback_Print method now abbreviates long sequences of repeated lines as "[Previous line repeated {count} more times]". (Contributed by Emanuel Barry in bpo-26823.)
  • The new PyErr_SetImportErrorSubclass() function allows for specifying a subclass of ImportError to raise. (Contributed by Eric Snow in bpo-15767.)
  • The new PyErr_ResourceWarning() function can be used to generate a ResourceWarning providing the source of the resource allocation. (Contributed by Victor Stinner in bpo-26567.)
  • The new PyOS_FSPath() function returns the file system representation of a path-like object. (Contributed by Brett Cannon in bpo-27186.)
  • The PyUnicode_FSConverter() and PyUnicode_FSDecoder() functions will now accept path-like objects.

Other Improvements

  • When --version (short form: -V) is supplied twice, Python prints sys.version for detailed information.

    $ ./python -VV
    Python 3.6.0b4+ (3.6:223967b49e49+, Nov 21 2016, 20:55:04)
    [GCC 4.2.1 Compatible Apple LLVM 8.0.0 (clang-800.0.42.1)]


New Keywords

async and await are not recommended to be used as variable, class, function or module names. Introduced by PEP 492 in Python 3.5, they will become proper keywords in Python 3.7. Starting in Python 3.6, the use of async or await as names will generate a DeprecationWarning.

Deprecated Python behavior

Raising the StopIteration exception inside a generator will now generate a DeprecationWarning, and will trigger a RuntimeError in Python 3.7. See PEP 479: Change StopIteration handling inside generators for details.

The __aiter__() method is now expected to return an asynchronous iterator directly instead of returning an awaitable as previously. Doing the former will trigger a DeprecationWarning. Backward compatibility will be removed in Python 3.7. (Contributed by Yury Selivanov in bpo-27243.)

A backslash-character pair that is not a valid escape sequence now generates a DeprecationWarning. Although this will eventually become a SyntaxError, that will not be for several Python releases. (Contributed by Emanuel Barry in bpo-27364.)

When performing a relative import, falling back on __name__ and __path__ from the calling module when __spec__ or __package__ are not defined now raises an ImportWarning. (Contributed by Rose Ames in bpo-25791.)

Deprecated Python modules, functions and methods


The asynchat has been deprecated in favor of asyncio. (Contributed by Mariatta in bpo-25002.)


The asyncore has been deprecated in favor of asyncio. (Contributed by Mariatta in bpo-25002.)


Unlike other dbm implementations, the dbm.dumb module creates databases with the 'rw' mode and allows modifying the database opened with the 'r' mode. This behavior is now deprecated and will be removed in 3.8. (Contributed by Serhiy Storchaka in bpo-21708.)


The undocumented extra_path argument to the Distribution constructor is now considered deprecated and will raise a warning if set. Support for this parameter will be removed in a future Python release. See bpo-27919 for details.


The support of non-integer arguments in getgrgid() has been deprecated. (Contributed by Serhiy Storchaka in bpo-26129.)


The importlib.machinery.SourceFileLoader.load_module() and importlib.machinery.SourcelessFileLoader.load_module() methods are now deprecated. They were the only remaining implementations of importlib.abc.Loader.load_module() in importlib that had not been deprecated in previous versions of Python in favour of importlib.abc.Loader.exec_module().

The importlib.machinery.WindowsRegistryFinder class is now deprecated. As of 3.6.0, it is still added to sys.meta_path by default (on Windows), but this may change in future releases.


Undocumented support of general bytes-like objects as paths in os functions, compile() and similar functions is now deprecated. (Contributed by Serhiy Storchaka in bpo-25791 and bpo-26754.)


Support for inline flags (?letters) in the middle of the regular expression has been deprecated and will be removed in a future Python version. Flags at the start of a regular expression are still allowed. (Contributed by Serhiy Storchaka in bpo-22493.)


OpenSSL 0.9.8, 1.0.0 and 1.0.1 are deprecated and no longer supported. In the future the ssl module will require at least OpenSSL 1.0.2 or 1.1.0.

SSL-related arguments like certfile, keyfile and check_hostname in ftplib, http.client, imaplib, poplib, and smtplib have been deprecated in favor of context. (Contributed by Christian Heimes in bpo-28022.)

A couple of protocols and functions of the ssl module are now deprecated. Some features will no longer be available in future versions of OpenSSL. Other features are deprecated in favor of a different API. (Contributed by Christian Heimes in bpo-28022 and bpo-26470.)


The tkinter.tix module is now deprecated. tkinter users should use tkinter.ttk instead.


The pyvenv script has been deprecated in favour of python3 -m venv. This prevents confusion as to what Python interpreter pyvenv is connected to and thus what Python interpreter will be used by the virtual environment. (Contributed by Brett Cannon in bpo-25154.)

Deprecated functions and types of the C API

Undocumented functions PyUnicode_AsEncodedObject(), PyUnicode_AsDecodedObject(), PyUnicode_AsEncodedUnicode() and PyUnicode_AsDecodedUnicode() are deprecated now. Use the generic codec based API instead.

Deprecated Build Options

The --with-system-ffi configure flag is now on by default on non-macOS UNIX platforms. It may be disabled by using --without-system-ffi, but using the flag is deprecated and will not be accepted in Python 3.7. macOS is unaffected by this change. Note that many OS distributors already use the --with-system-ffi flag when building their system Python.


API and Feature Removals

  • Unknown escapes consisting of '\' and an ASCII letter in regular expressions will now cause an error. In replacement templates for re.sub() they are still allowed, but deprecated. The re.LOCALE flag can now only be used with binary patterns.
  • inspect.getmoduleinfo() was removed (was deprecated since CPython 3.3). inspect.getmodulename() should be used for obtaining the module name for a given path. (Contributed by Yury Selivanov in bpo-13248.)
  • traceback.Ignore class and traceback.usage, traceback.modname, traceback.fullmodname, traceback.find_lines_from_code, traceback.find_lines, traceback.find_strings, traceback.find_executable_lines methods were removed from the traceback module. They were undocumented methods deprecated since Python 3.2 and equivalent functionality is available from private methods.
  • The tk_menuBar() and tk_bindForTraversal() dummy methods in tkinter widget classes were removed (corresponding Tk commands were obsolete since Tk 4.0).
  • The open() method of the zipfile.ZipFile class no longer supports the 'U' mode (was deprecated since Python 3.4). Use io.TextIOWrapper for reading compressed text files in universal newlines mode.
  • The undocumented IN, CDROM, DLFCN, TYPES, CDIO, and STROPTS modules have been removed. They had been available in the platform specific Lib/plat-*/ directories, but were chronically out of date, inconsistently available across platforms, and unmaintained. The script that created these modules is still available in the source distribution at Tools/scripts/h2py.py.
  • The deprecated asynchat.fifo class has been removed.

Porting to Python 3.6

This section lists previously described changes and other bugfixes that may require changes to your code.

Changes in ‘python’ Command Behavior

  • The output of a special Python build with defined COUNT_ALLOCS, SHOW_ALLOC_COUNT or SHOW_TRACK_COUNT macros is now off by default. It can be re-enabled using the -X showalloccount option. It now outputs to stderr instead of stdout. (Contributed by Serhiy Storchaka in bpo-23034.)

Changes in the Python API

  • open() will no longer allow combining the 'U' mode flag with '+'. (Contributed by Jeff Balogh and John O’Connor in bpo-2091.)

  • sqlite3 no longer implicitly commits an open transaction before DDL statements.

  • On Linux, os.urandom() now blocks until the system urandom entropy pool is initialized to increase the security.

  • When importlib.abc.Loader.exec_module() is defined, importlib.abc.Loader.create_module() must also be defined.

  • PyErr_SetImportError() now sets TypeError when its msg argument is not set. Previously only NULL was returned.

  • The format of the co_lnotab attribute of code objects changed to support a negative line number delta. By default, Python does not emit bytecode with a negative line number delta. Functions using frame.f_lineno, PyFrame_GetLineNumber() or PyCode_Addr2Line() are not affected. Functions directly decoding co_lnotab should be updated to use a signed 8-bit integer type for the line number delta, but this is only required to support applications using a negative line number delta. See Objects/lnotab_notes.txt for the co_lnotab format and how to decode it, and see the PEP 511 for the rationale.

  • The functions in the compileall module now return booleans instead of 1 or 0 to represent success or failure, respectively. Thanks to booleans being a subclass of integers, this should only be an issue if you were doing identity checks for 1 or 0. See bpo-25768.

  • Reading the port attribute of urllib.parse.urlsplit() and urlparse() results now raises ValueError for out-of-range values, rather than returning None. See bpo-20059.

  • The imp module now raises a DeprecationWarning instead of PendingDeprecationWarning.

  • The following modules have had missing APIs added to their __all__ attributes to match the documented APIs: calendar, cgi, csv, ElementTree, enum, fileinput, ftplib, logging, mailbox, mimetypes, optparse, plistlib, smtpd, subprocess, tarfile, threading and wave. This means they will export new symbols when import * is used. (Contributed by Joel Taddei and Jacek Kołodziej in bpo-23883.)

  • When performing a relative import, if __package__ does not compare equal to __spec__.parent then ImportWarning is raised. (Contributed by Brett Cannon in bpo-25791.)

  • When a relative import is performed and no parent package is known, then ImportError will be raised. Previously, SystemError could be raised. (Contributed by Brett Cannon in bpo-18018.)

  • Servers based on the socketserver module, including those defined in http.server, xmlrpc.server and wsgiref.simple_server, now only catch exceptions derived from Exception. Therefore if a request handler raises an exception like SystemExit or KeyboardInterrupt, handle_error() is no longer called, and the exception will stop a single-threaded server. (Contributed by Martin Panter in bpo-23430.)

  • spwd.getspnam() now raises a PermissionError instead of KeyError if the user doesn’t have privileges.

  • The socket.socket.close() method now raises an exception if an error (e.g. EBADF) was reported by the underlying system call. (Contributed by Martin Panter in bpo-26685.)

  • The decode_data argument for the smtpd.SMTPChannel and smtpd.SMTPServer constructors is now False by default. This means that the argument passed to process_message() is now a bytes object by default, and process_message() will be passed keyword arguments. Code that has already been updated in accordance with the deprecation warning generated by 3.5 will not be affected.

  • All optional arguments of the dump(), dumps(), load() and loads() functions and JSONEncoder and JSONDecoder class constructors in the json module are now keyword-only. (Contributed by Serhiy Storchaka in bpo-18726.)

  • Subclasses of type which don’t override type.__new__ may no longer use the one-argument form to get the type of an object.

  • As part of PEP 487, the handling of keyword arguments passed to type (other than the metaclass hint, metaclass) is now consistently delegated to object.__init_subclass__(). This means that type.__new__() and type.__init__() both now accept arbitrary keyword arguments, but object.__init_subclass__() (which is called from type.__new__()) will reject them by default. Custom metaclasses accepting additional keyword arguments will need to adjust their calls to type.__new__() (whether direct or via super) accordingly.

  • In distutils.command.sdist.sdist, the default_format attribute has been removed and is no longer honored. Instead, the gzipped tarfile format is the default on all platforms and no platform-specific selection is made. In environments where distributions are built on Windows and zip distributions are required, configure the project with a setup.cfg file containing the following:


    This behavior has also been backported to earlier Python versions by Setuptools 26.0.0.

  • In the urllib.request module and the http.client.HTTPConnection.request() method, if no Content-Length header field has been specified and the request body is a file object, it is now sent with HTTP 1.1 chunked encoding. If a file object has to be sent to a HTTP 1.0 server, the Content-Length value now has to be specified by the caller. (Contributed by Demian Brecht and Rolf Krahl with tweaks from Martin Panter in bpo-12319.)

  • The DictReader now returns rows of type OrderedDict. (Contributed by Steve Holden in bpo-27842.)

  • The crypt.METHOD_CRYPT will no longer be added to crypt.methods if unsupported by the platform. (Contributed by Victor Stinner in bpo-25287.)

  • The verbose and rename arguments for namedtuple() are now keyword-only. (Contributed by Raymond Hettinger in bpo-25628.)

  • On Linux, ctypes.util.find_library() now looks in LD_LIBRARY_PATH for shared libraries. (Contributed by Vinay Sajip in bpo-9998.)

  • The imaplib.IMAP4 class now handles flags containing the ']' character in messages sent from the server to improve real-world compatibility. (Contributed by Lita Cho in bpo-21815.)

  • The mmap.write() function now returns the number of bytes written like other write methods. (Contributed by Jakub Stasiak in bpo-26335.)

  • The pkgutil.iter_modules() and pkgutil.walk_packages() functions now return ModuleInfo named tuples. (Contributed by Ramchandra Apte in bpo-17211.)

  • re.sub() now raises an error for invalid numerical group references in replacement templates even if the pattern is not found in the string. The error message for invalid group references now includes the group index and the position of the reference. (Contributed by SilentGhost, Serhiy Storchaka in bpo-25953.)

  • zipfile.ZipFile will now raise NotImplementedError for unrecognized compression values. Previously a plain RuntimeError was raised. Additionally, calling ZipFile methods on a closed ZipFile or calling the write() method on a ZipFile created with mode 'r' will raise a ValueError. Previously, a RuntimeError was raised in those scenarios.

  • when custom metaclasses are combined with zero-argument super() or direct references from methods to the implicit __class__ closure variable, the implicit __classcell__ namespace entry must now be passed up to type.__new__ for initialisation. Failing to do so will result in a DeprecationWarning in 3.6 and a RuntimeWarning in the future.

Changes in the C API

  • The PyMem_Malloc() allocator family now uses the pymalloc allocator rather than the system malloc(). Applications calling PyMem_Malloc() without holding the GIL can now crash. Set the PYTHONMALLOC environment variable to debug to validate the usage of memory allocators in your application. See bpo-26249.
  • Py_Exit() (and the main interpreter) now override the exit status with 120 if flushing buffered data failed. See bpo-5319.

CPython bytecode changes

There have been several major changes to the bytecode in Python 3.6.