Python 3.12 有什麼新功能

編輯者:

Adam Turner

本文介紹了 Python 3.12 與 3.11 相比多了哪些新功能。 Python 3.12 於 2023 年 10 月 2 日發布。完整詳情請見 changelog。

也參考

PEP 693 -- Python 3.12 發佈時程

發布重點摘要

Python 3.12 is a stable release of the Python programming language, with a mix of changes to the language and the standard library. The library changes focus on cleaning up deprecated APIs, usability, and correctness. Of note, the distutils package has been removed from the standard library. Filesystem support in os and pathlib has seen a number of improvements, and several modules have better performance.

The language changes focus on usability, as f-strings have had many limitations removed and 'Did you mean ...' suggestions continue to improve. The new type parameter syntax and type statement improve ergonomics for using generic types and type aliases with static type checkers.

This article doesn't attempt to provide a complete specification of all new features, but instead gives a convenient overview. For full details, you should refer to the documentation, such as the Library Reference and Language Reference. If you want to understand the complete implementation and design rationale for a change, refer to the PEP for a particular new feature; but note that PEPs usually are not kept up-to-date once a feature has been fully implemented.

新增語法特性：

新增語法特性：

直譯器改進：

Python 資料模型改進：

標準函式庫中的顯著改進

安全性改進：

  • Replace the builtin hashlib implementations of SHA1, SHA3, SHA2-384, SHA2-512, and MD5 with formally verified code from the HACL* project. These builtin implementations remain as fallbacks that are only used when OpenSSL does not provide them.

C API 改進：

CPython 實作改進：

  • PEP 709, comprehension inlining

  • CPython support for the Linux perf profiler

  • Implement stack overflow protection on supported platforms

新增型別特性：

重要的棄用、刪除或限制：

  • PEP 623: Remove wstr from Unicode objects in Python's C API, reducing the size of every str object by at least 8 bytes.

  • PEP 632: Remove the distutils package. See the migration guide for advice replacing the APIs it provided. The third-party Setuptools package continues to provide distutils, if you still require it in Python 3.12 and beyond.

  • gh-95299: Do not pre-install setuptools in virtual environments created with venv. This means that distutils, setuptools, pkg_resources, and easy_install will no longer available by default; to access these run pip install setuptools in the activated virtual environment.

  • asynchat、asyncore 和 imp 模組以及幾個 unittest.TestCase 的方法別名已被刪除。

新增功能

PEP 695：型別參數語法

Generic classes and functions under PEP 484 were declared using a verbose syntax that left the scope of type parameters unclear and required explicit declarations of variance.

PEP 695 introduces a new, more compact and explicit way to create generic classes and functions:

def max[T](args: Iterable[T]) -> T:
    ...

class list[T]:
    def __getitem__(self, index: int, /) -> T:
        ...

    def append(self, element: T) -> None:
        ...

In addition, the PEP introduces a new way to declare type aliases using the type statement, which creates an instance of TypeAliasType:

type Point = tuple[float, float]

型別別名也可以是泛型：

type Point[T] = tuple[T, T]

The new syntax allows declaring TypeVarTuple and ParamSpec parameters, as well as TypeVar parameters with bounds or constraints:

type IntFunc[**P] = Callable[P, int]  # ParamSpec
type LabeledTuple[*Ts] = tuple[str, *Ts]  # TypeVarTuple
type HashableSequence[T: Hashable] = Sequence[T]  # TypeVar with bound
type IntOrStrSequence[T: (int, str)] = Sequence[T]  # TypeVar with constraints

The value of type aliases and the bound and constraints of type variables created through this syntax are evaluated only on demand (see lazy evaluation). This means type aliases are able to refer to other types defined later in the file.

Type parameters declared through a type parameter list are visible within the scope of the declaration and any nested scopes, but not in the outer scope. For example, they can be used in the type annotations for the methods of a generic class or in the class body. However, they cannot be used in the module scope after the class is defined. See Type parameter lists for a detailed description of the runtime semantics of type parameters.

In order to support these scoping semantics, a new kind of scope is introduced, the annotation scope. Annotation scopes behave for the most part like function scopes, but interact differently with enclosing class scopes. In Python 3.13, annotations will also be evaluated in annotation scopes.

詳情請見 PEP 695。

(PEP written by Eric Traut. Implementation by Jelle Zijlstra, Eric Traut, and others in gh-103764.)

PEP 701: Syntactic formalization of f-strings

PEP 701 lifts some restrictions on the usage of f-strings. Expression components inside f-strings can now be any valid Python expression, including strings reusing the same quote as the containing f-string, multi-line expressions, comments, backslashes, and unicode escape sequences. Let's cover these in detail:

  • Quote reuse: in Python 3.11, reusing the same quotes as the enclosing f-string raises a SyntaxError, forcing the user to either use other available quotes (like using double quotes or triple quotes if the f-string uses single quotes). In Python 3.12, you can now do things like this:

    >>> songs = ['Take me back to Eden', 'Alkaline', 'Ascensionism']
>>> f"This is the playlist: {", ".join(songs)}"
'This is the playlist: Take me back to Eden, Alkaline, Ascensionism'

    Note that before this change there was no explicit limit in how f-strings can be nested, but the fact that string quotes cannot be reused inside the expression component of f-strings made it impossible to nest f-strings arbitrarily. In fact, this is the most nested f-string that could be written:

    >>> f"""{f'''{f'{f"{1+1}"}'}'''}"""
'2'

    As now f-strings can contain any valid Python expression inside expression components, it is now possible to nest f-strings arbitrarily:

    >>> f"{f"{f"{f"{f"{f"{1+1}"}"}"}"}"}"
'2'

  • Multi-line expressions and comments: In Python 3.11, f-string expressions must be defined in a single line, even if the expression within the f-string could normally span multiple lines (like literal lists being defined over multiple lines), making them harder to read. In Python 3.12 you can now define f-strings spanning multiple lines, and add inline comments:

    >>> f"This is the playlist: {", ".join([
...     'Take me back to Eden',  # My, my, those eyes like fire
...     'Alkaline',              # Not acid nor alkaline
...     'Ascensionism'           # Take to the broken skies at last
... ])}"
'This is the playlist: Take me back to Eden, Alkaline, Ascensionism'

  • Backslashes and unicode characters: before Python 3.12 f-string expressions couldn't contain any \ character. This also affected unicode escape sequences (such as \N{snowman}) as these contain the \N part that previously could not be part of expression components of f-strings. Now, you can define expressions like this:

    >>> print(f"This is the playlist: {"\n".join(songs)}")
This is the playlist: Take me back to Eden
Alkaline
Ascensionism
>>> print(f"This is the playlist: {"\N{BLACK HEART SUIT}".join(songs)}")
This is the playlist: Take me back to Eden♥Alkaline♥Ascensionism

詳情請見 PEP 701。

As a positive side-effect of how this feature has been implemented (by parsing f-strings with the PEG parser), now error messages for f-strings are more precise and include the exact location of the error. For example, in Python 3.11, the following f-string raises a SyntaxError:

>>> my_string = f"{x z y}" + f"{1 + 1}"
  File "<stdin>", line 1
    (x z y)
     ^^^
SyntaxError: f-string: invalid syntax. Perhaps you forgot a comma?

but the error message doesn't include the exact location of the error within the line and also has the expression artificially surrounded by parentheses. In Python 3.12, as f-strings are parsed with the PEG parser, error messages can be more precise and show the entire line:

>>> my_string = f"{x z y}" + f"{1 + 1}"
  File "<stdin>", line 1
    my_string = f"{x z y}" + f"{1 + 1}"
                   ^^^
SyntaxError: invalid syntax. Perhaps you forgot a comma?

(Contributed by Pablo Galindo, Batuhan Taskaya, Lysandros Nikolaou, Cristián Maureira-Fredes and Marta Gómez in gh-102856. PEP written by Pablo Galindo, Batuhan Taskaya, Lysandros Nikolaou and Marta Gómez).

PEP 684: A Per-Interpreter GIL

PEP 684 introduces a per-interpreter GIL, so that sub-interpreters may now be created with a unique GIL per interpreter. This allows Python programs to take full advantage of multiple CPU cores. This is currently only available through the C-API, though a Python API is anticipated for 3.13.

Use the new Py_NewInterpreterFromConfig() function to create an interpreter with its own GIL:

PyInterpreterConfig config = {
    .check_multi_interp_extensions = 1,
    .gil = PyInterpreterConfig_OWN_GIL,
};
PyThreadState *tstate = NULL;
PyStatus status = Py_NewInterpreterFromConfig(&tstate, &config);
if (PyStatus_Exception(status)) {
    return -1;
}
/* The new interpreter is now active in the current thread. */

For further examples how to use the C-API for sub-interpreters with a per-interpreter GIL, see Modules/_xxsubinterpretersmodule.c.

（由 Eric Snow 於 gh-104210 等貢獻。）

PEP 669: Low impact monitoring for CPython

PEP 669 defines a new API for profilers, debuggers, and other tools to monitor events in CPython. It covers a wide range of events, including calls, returns, lines, exceptions, jumps, and more. This means that you only pay for what you use, providing support for near-zero overhead debuggers and coverage tools. See sys.monitoring for details.

（由 Mark Shannon 於 gh-103082 中貢獻。）

PEP 688: Making the buffer protocol accessible in Python

PEP 688 introduces a way to use the buffer protocol from Python code. Classes that implement the __buffer__() method are now usable as buffer types.

The new collections.abc.Buffer ABC provides a standard way to represent buffer objects, for example in type annotations. The new inspect.BufferFlags enum represents the flags that can be used to customize buffer creation. (Contributed by Jelle Zijlstra in gh-102500.)

PEP 709: Comprehension inlining

Dictionary, list, and set comprehensions are now inlined, rather than creating a new single-use function object for each execution of the comprehension. This speeds up execution of a comprehension by up to two times. See PEP 709 for further details.

Comprehension iteration variables remain isolated and don't overwrite a variable of the same name in the outer scope, nor are they visible after the comprehension. Inlining does result in a few visible behavior changes:

  • There is no longer a separate frame for the comprehension in tracebacks, and tracing/profiling no longer shows the comprehension as a function call.

  • The symtable module will no longer produce child symbol tables for each comprehension; instead, the comprehension's locals will be included in the parent function's symbol table.

  • Calling locals() inside a comprehension now includes variables from outside the comprehension, and no longer includes the synthetic .0 variable for the comprehension "argument".

  • A comprehension iterating directly over locals() (e.g. [k for k in locals()]) may see "RuntimeError: dictionary changed size during iteration" when run under tracing (e.g. code coverage measurement). This is the same behavior already seen in e.g. for k in locals():. To avoid the error, first create a list of keys to iterate over: keys = list(locals()); [k for k in keys].

（由 Carl Meyer 和 Vladimir Matveev 於 PEP 709 中貢獻。）

Improved Error Messages

  • Modules from the standard library are now potentially suggested as part of the error messages displayed by the interpreter when a NameError is raised to the top level. (Contributed by Pablo Galindo in gh-98254.)

    >>> sys.version_info
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
NameError: name 'sys' is not defined. Did you forget to import 'sys'?

  • Improve the error suggestion for NameError exceptions for instances. Now if a NameError is raised in a method and the instance has an attribute that's exactly equal to the name in the exception, the suggestion will include self.<NAME> instead of the closest match in the method scope. (Contributed by Pablo Galindo in gh-99139.)

    >>> class A:
...    def __init__(self):
...        self.blech = 1
...
...    def foo(self):
...        somethin = blech
...
>>> A().foo()
Traceback (most recent call last):
  File "<stdin>", line 1
    somethin = blech
               ^^^^^
NameError: name 'blech' is not defined. Did you mean: 'self.blech'?

  • Improve the SyntaxError error message when the user types import x from y instead of from y import x. (Contributed by Pablo Galindo in gh-98931.)

    >>> import a.y.z from b.y.z
Traceback (most recent call last):
  File "<stdin>", line 1
    import a.y.z from b.y.z
    ^^^^^^^^^^^^^^^^^^^^^^^
SyntaxError: Did you mean to use 'from ... import ...' instead?

  • ImportError exceptions raised from failed from <module> import <name> statements now include suggestions for the value of <name> based on the available names in <module>. (Contributed by Pablo Galindo in gh-91058.)

    >>> from collections import chainmap
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: cannot import name 'chainmap' from 'collections'. Did you mean: 'ChainMap'?

其他語言更動

  • The parser now raises SyntaxError when parsing source code containing null bytes. (Contributed by Pablo Galindo in gh-96670.)

  • A backslash-character pair that is not a valid escape sequence now generates a SyntaxWarning, instead of DeprecationWarning. For example, re.compile("\d+\.\d+") now emits a SyntaxWarning ("\d" is an invalid escape sequence, use raw strings for regular expression: re.compile(r"\d+\.\d+")). In a future Python version, SyntaxError will eventually be raised, instead of SyntaxWarning. (Contributed by Victor Stinner in gh-98401.)

  • Octal escapes with value larger than 0o377 (ex: "\477"), deprecated in Python 3.11, now produce a SyntaxWarning, instead of DeprecationWarning. In a future Python version they will be eventually a SyntaxError. (Contributed by Victor Stinner in gh-98401.)

  • Variables used in the target part of comprehensions that are not stored to can now be used in assignment expressions (:=). For example, in [(b := 1) for a, b.prop in some_iter], the assignment to b is now allowed. Note that assigning to variables stored to in the target part of comprehensions (like a) is still disallowed, as per PEP 572. (Contributed by Nikita Sobolev in gh-100581.)

  • Exceptions raised in a class or type's __set_name__ method are no longer wrapped by a RuntimeError. Context information is added to the exception as a PEP 678 note. (Contributed by Irit Katriel in gh-77757.)

  • When a try-except* construct handles the entire ExceptionGroup and raises one other exception, that exception is no longer wrapped in an ExceptionGroup. Also changed in version 3.11.4. (Contributed by Irit Katriel in gh-103590.)

  • The Garbage Collector now runs only on the eval breaker mechanism of the Python bytecode evaluation loop instead of object allocations. The GC can also run when PyErr_CheckSignals() is called so C extensions that need to run for a long time without executing any Python code also have a chance to execute the GC periodically. (Contributed by Pablo Galindo in gh-97922.)

  • All builtin and extension callables expecting boolean parameters now accept arguments of any type instead of just bool and int. (Contributed by Serhiy Storchaka in gh-60203.)

  • memoryview now supports the half-float type (the "e" format code). (Contributed by Donghee Na and Antoine Pitrou in gh-90751.)

  • slice objects are now hashable, allowing them to be used as dict keys and set items. (Contributed by Will Bradshaw, Furkan Onder, and Raymond Hettinger in gh-101264.)

  • sum() now uses Neumaier summation to improve accuracy and commutativity when summing floats or mixed ints and floats. (Contributed by Raymond Hettinger in gh-100425.)

  • ast.parse() now raises SyntaxError instead of ValueError when parsing source code containing null bytes. (Contributed by Pablo Galindo in gh-96670.)

  • The extraction methods in tarfile, and shutil.unpack_archive(), have a new a filter argument that allows limiting tar features than may be surprising or dangerous, such as creating files outside the destination directory. See tarfile extraction filters for details. In Python 3.14, the default will switch to 'data'. (Contributed by Petr Viktorin in PEP 706.)

  • types.MappingProxyType instances are now hashable if the underlying mapping is hashable. (Contributed by Serhiy Storchaka in gh-87995.)

  • Add support for the perf profiler through the new environment variable PYTHONPERFSUPPORT and command-line option -X perf, as well as the new sys.activate_stack_trampoline(), sys.deactivate_stack_trampoline(), and sys.is_stack_trampoline_active() functions. (Design by Pablo Galindo. Contributed by Pablo Galindo and Christian Heimes with contributions from Gregory P. Smith [Google] and Mark Shannon in gh-96123.)

新增模組

  • None.

改進的模組

array

asyncio

  • The performance of writing to sockets in asyncio has been significantly improved. asyncio now avoids unnecessary copying when writing to sockets and uses sendmsg() if the platform supports it. (Contributed by Kumar Aditya in gh-91166.)

  • Add asyncio.eager_task_factory() and asyncio.create_eager_task_factory() functions to allow opting an event loop in to eager task execution, making some use-cases 2x to 5x faster. (Contributed by Jacob Bower & Itamar Oren in gh-102853, gh-104140, and gh-104138)

  • On Linux, asyncio uses asyncio.PidfdChildWatcher by default if os.pidfd_open() is available and functional instead of asyncio.ThreadedChildWatcher. (Contributed by Kumar Aditya in gh-98024.)

  • The event loop now uses the best available child watcher for each platform (asyncio.PidfdChildWatcher if supported and asyncio.ThreadedChildWatcher otherwise), so manually configuring a child watcher is not recommended. (Contributed by Kumar Aditya in gh-94597.)

  • Add loop_factory parameter to asyncio.run() to allow specifying a custom event loop factory. (Contributed by Kumar Aditya in gh-99388.)

  • Add C implementation of asyncio.current_task() for 4x-6x speedup. (Contributed by Itamar Oren and Pranav Thulasiram Bhat in gh-100344.)

  • asyncio.iscoroutine() now returns False for generators as asyncio does not support legacy generator-based coroutines. (Contributed by Kumar Aditya in gh-102748.)

  • asyncio.wait() and asyncio.as_completed() now accepts generators yielding tasks. (Contributed by Kumar Aditya in gh-78530.)

calendar

csv

dis

  • Pseudo instruction opcodes (which are used by the compiler but do not appear in executable bytecode) are now exposed in the dis module. HAVE_ARGUMENT is still relevant to real opcodes, but it is not useful for pseudo instructions. Use the new dis.hasarg collection instead. (Contributed by Irit Katriel in gh-94216.)

  • Add the dis.hasexc collection to signify instructions that set an exception handler. (Contributed by Irit Katriel in gh-94216.)

fractions

importlib.resources

inspect

itertools

  • Add itertools.batched() for collecting into even-sized tuples where the last batch may be shorter than the rest. (Contributed by Raymond Hettinger in gh-98363.)

math

  • 新增 math.sumprod() 以計算乘積總和。（由 Raymond Hettinger 於 gh-100485 中貢獻。）

  • Extend math.nextafter() to include a steps argument for moving up or down multiple steps at a time. (Contributed by Matthias Goergens, Mark Dickinson, and Raymond Hettinger in gh-94906.)

os

  • Add os.PIDFD_NONBLOCK to open a file descriptor for a process with os.pidfd_open() in non-blocking mode. (Contributed by Kumar Aditya in gh-93312.)

  • os.DirEntry now includes an os.DirEntry.is_junction() method to check if the entry is a junction. (Contributed by Charles Machalow in gh-99547.)

  • Add os.listdrives(), os.listvolumes() and os.listmounts() functions on Windows for enumerating drives, volumes and mount points. (Contributed by Steve Dower in gh-102519.)

  • os.stat() and os.lstat() are now more accurate on Windows. The st_birthtime field will now be filled with the creation time of the file, and st_ctime is deprecated but still contains the creation time (but in the future will return the last metadata change, for consistency with other platforms). st_dev may be up to 64 bits and st_ino up to 128 bits depending on your file system, and st_rdev is always set to zero rather than incorrect values. Both functions may be significantly faster on newer releases of Windows. (Contributed by Steve Dower in gh-99726.)

os.path

pathlib

pdb

  • Add convenience variables to hold values temporarily for debug session and provide quick access to values like the current frame or the return value. (Contributed by Tian Gao in gh-103693.)

random

shutil

  • shutil.make_archive() now passes the root_dir argument to custom archivers which support it. In this case it no longer temporarily changes the current working directory of the process to root_dir to perform archiving. (Contributed by Serhiy Storchaka in gh-74696.)

  • shutil.rmtree() now accepts a new argument onexc which is an error handler like onerror but which expects an exception instance rather than a (typ, val, tb) triplet. onerror is deprecated. (Contributed by Irit Katriel in gh-102828.)

  • shutil.which() now consults the PATHEXT environment variable to find matches within PATH on Windows even when the given cmd includes a directory component. (Contributed by Charles Machalow in gh-103179.)

    shutil.which() will call NeedCurrentDirectoryForExePathW when querying for executables on Windows to determine if the current working directory should be prepended to the search path. (Contributed by Charles Machalow in gh-103179.)

    shutil.which() will return a path matching the cmd with a component from PATHEXT prior to a direct match elsewhere in the search path on Windows. (Contributed by Charles Machalow in gh-103179.)

sqlite3

statistics

  • Extend statistics.correlation() to include as a ranked method for computing the Spearman correlation of ranked data. (Contributed by Raymond Hettinger in gh-95861.)

sys

tempfile

threading

tkinter

  • tkinter.Canvas.coords() now flattens its arguments. It now accepts not only coordinates as separate arguments (x1, y1, x2, y2, ...) and a sequence of coordinates ([x1, y1, x2, y2, ...]), but also coordinates grouped in pairs ((x1, y1), (x2, y2), ... and [(x1, y1), (x2, y2), ...]), like create_*() methods. (Contributed by Serhiy Storchaka in gh-94473.)

tokenize

types

typing

  • isinstance() checks against runtime-checkable protocols now use inspect.getattr_static() rather than hasattr() to lookup whether attributes exist. This means that descriptors and __getattr__() methods are no longer unexpectedly evaluated during isinstance() checks against runtime-checkable protocols. However, it may also mean that some objects which used to be considered instances of a runtime-checkable protocol may no longer be considered instances of that protocol on Python 3.12+, and vice versa. Most users are unlikely to be affected by this change. (Contributed by Alex Waygood in gh-102433.)

  • The members of a runtime-checkable protocol are now considered "frozen" at runtime as soon as the class has been created. Monkey-patching attributes onto a runtime-checkable protocol will still work, but will have no impact on isinstance() checks comparing objects to the protocol. For example:

    >>> from typing import Protocol, runtime_checkable
>>> @runtime_checkable
... class HasX(Protocol):
...     x = 1
...
>>> class Foo: ...
...
>>> f = Foo()
>>> isinstance(f, HasX)
False
>>> f.x = 1
>>> isinstance(f, HasX)
True
>>> HasX.y = 2
>>> isinstance(f, HasX)  # unchanged, even though HasX now also has a "y" attribute
True

    This change was made in order to speed up isinstance() checks against runtime-checkable protocols.

  • The performance profile of isinstance() checks against runtime-checkable protocols has changed significantly. Most isinstance() checks against protocols with only a few members should be at least 2x faster than in 3.11, and some may be 20x faster or more. However, isinstance() checks against protocols with many members may be slower than in Python 3.11. (Contributed by Alex Waygood in gh-74690 and gh-103193.)

  • All typing.TypedDict and typing.NamedTuple classes now have the __orig_bases__ attribute. (Contributed by Adrian Garcia Badaracco in gh-103699.)

  • 新增 frozen_default 參數至 typing.dataclass_transform()。（由 Erik De Bonte 於 gh-99957 中貢獻。）

unicodedata

  • The Unicode database has been updated to version 15.0.0. (Contributed by Benjamin Peterson in gh-96734).

unittest

新增 --durations 命令列選項，顯示 N 個最慢的測試案例：

python3 -m unittest --durations=3 lib.tests.test_threading
.....
Slowest test durations
----------------------------------------------------------------------
1.210s     test_timeout (Lib.test.test_threading.BarrierTests)
1.003s     test_default_timeout (Lib.test.test_threading.BarrierTests)
0.518s     test_timeout (Lib.test.test_threading.EventTests)

(0.000 durations hidden.  Use -v to show these durations.)
----------------------------------------------------------------------
Ran 158 tests in 9.869s

OK (skipped=3)

（由 Giampaolo Rodola 於 gh-48330 中貢獻。）

uuid

最佳化

  • Remove wstr and wstr_length members from Unicode objects. It reduces object size by 8 or 16 bytes on 64bit platform. (PEP 623) (Contributed by Inada Naoki in gh-92536.)

  • Add experimental support for using the BOLT binary optimizer in the build process, which improves performance by 1-5%. (Contributed by Kevin Modzelewski in gh-90536 and tuned by Donghee Na in gh-101525)

  • Speed up the regular expression substitution (functions re.sub() and re.subn() and corresponding re.Pattern methods) for replacement strings containing group references by 2--3 times. (Contributed by Serhiy Storchaka in gh-91524.)

  • Speed up asyncio.Task creation by deferring expensive string formatting. (Contributed by Itamar Oren in gh-103793.)

  • The tokenize.tokenize() and tokenize.generate_tokens() functions are up to 64% faster as a side effect of the changes required to cover PEP 701 in the tokenize module. (Contributed by Marta Gómez Macías and Pablo Galindo in gh-102856.)

  • Speed up super() method calls and attribute loads via the new LOAD_SUPER_ATTR instruction. (Contributed by Carl Meyer and Vladimir Matveev in gh-103497.)

CPython 位元組碼變更

Demos and Tools

  • Remove the Tools/demo/ directory which contained old demo scripts. A copy can be found in the old-demos project. (Contributed by Victor Stinner in gh-97681.)

  • Remove outdated example scripts of the Tools/scripts/ directory. A copy can be found in the old-demos project. (Contributed by Victor Stinner in gh-97669.)

已棄用

Pending removal in Python 3.13

模組（請見 PEP 594）：

  • aifc

  • audioop

  • cgi

  • cgitb

  • chunk

  • crypt

  • imghdr

  • mailcap

  • msilib

  • nis

  • nntplib

  • ossaudiodev

  • pipes

  • sndhdr

  • spwd

  • sunau

  • telnetlib

  • uu

  • xdrlib

其他模組：

API：

Pending removal in Python 3.14

Pending removal in Python 3.15

  • 引入系統 (import system)：

    • 在模組上設定 __cached__ 而沒有設定 __spec__.cached 的做法已被棄用。在 Python 3.15 中，引入系統或標準函式庫將不再設定或考慮 __cached__。(gh-97879)

    • 在模組上設定 __package__ 而沒有設定 __spec__.parent 的做法已被棄用。在 Python 3.15 中，引入系統或標準函式庫將不再設定或考慮 __package__。(gh-97879)

  • ctypes：

    • 自 Python 3.13 起，未記錄的 ctypes.SetPointerType() 函式已被棄用。

  • http.server：

    • 過時且很少使用的 CGIHTTPRequestHandler 自 Python 3.13 起已被棄用。不存在直接的替代。任何東西都比 CGI 更好地將 Web 伺服器與請求處理程序介接起來。

    • 自 Python 3.13 起，python -m http.server 命令列介面的 --cgi 旗標已被棄用。

  • locale：

  • pathlib：

  • platform：

    • 自 Python 3.13 起，java_ver() 已被棄用。此函式僅對 Jython 支援有用，具有令人困惑的 API，基本上未經測試。

  • threading：

    • RLock() 在 Python 3.15 中將不接受任何引數。自 Python 3.14 起，傳遞任何引數的用法已被棄用，因為 Python 版本不允許任何引數，但 C 版本允許任意數量的位置或關鍵字引數，並忽略每個引數。

  • types：

  • typing：

    • The undocumented keyword argument syntax for creating NamedTuple classes (for example, Point = NamedTuple("Point", x=int, y=int)) has been deprecated since Python 3.13. Use the class-based syntax or the functional syntax instead.

    • 自 Python 3.13 起，typing.no_type_check_decorator() 裝飾器函式已被棄用。在 typing 模組中使用了八年之後，它尚未得到任何主要型別檢查器的支援。

  • wave：

Python 3.16 中待移除的項目

  • 引入系統 (import system)：

    • 在模組上設定 __loader__ 而沒有設定 __spec__.loader 的做法將於 Python 3.16 被棄用。在 Python 3.16 中，引入系統或標準函式庫將不再設定或考慮 __loader__。

  • array：

    • 自 Python 3.3 起，'u' 格式碼 (wchar_t) 在文件中已被棄用，自 Python 3.13 起在 runtime 已被棄用。請使用 'w' 格式碼 (Py_UCS4) 來取代 Unicode 字元。

  • asyncio：

  • builtins：

    • 自 Python 3.12 起，布林型別的位元反轉 ~True 或 ~False 已被棄用，因為它會產生不預期且不直觀的結果（-2 和 -1）。使用 not x 代替布林值的邏輯否定。在極少數情況下，你需要對底層的整數進行位元反轉，請明確轉換為 ~int(x) (~int(x))。

  • shutil：

    • 自 Python 3.14 起，ExecError 例外已被棄用。自 Python 3.4 以來，它尚未被 shutil 中的任何函式使用，現在是 RuntimeError 的別名。

  • symtable：

  • sys：

  • tarfile：

    • 自 Python 3.13 起，未以文件記錄和未被使用的 TarFile.tarfile 屬性已被棄用。

Pending removal in future versions

以下 API 將在未來被移除，雖然目前尚未安排移除日期。

  • argparse:

    • Nesting argument groups and nesting mutually exclusive groups are deprecated.

    • Passing the undocumented keyword argument prefix_chars to add_argument_group() is now deprecated.

    • The argparse.FileType type converter is deprecated.

  • array 的 'u' 格式碼 (gh-57281)

  • builtins：

    • bool(NotImplemented)。

    • 產生器：throw(type, exc, tb) 和 athrow(type, exc, tb) 簽名已被棄用：請改用 throw(exc) 和 athrow(exc)，為單引數簽名。

    • 目前 Python 接受數值字面值後面立即接關鍵字，例如 0in x、1or x、0if 1else 2。它讓表達式模糊且容易混淆，如 [0x1for x in y]（可以解釋為 [0x1 for x in y] 或 [0x1f or x in y]）。如果數值字面值後立即接 and、else、for、if、in、is 和 or 之一的關鍵字，則會引發語法警告。在未來版本中，它將被更改為語法錯誤。(gh-87999)

    • __index__() 和 __int__() 方法回傳非 int 型別的支援：這些方法將需要回傳 int 的嚴格子類別實例。

    • 回傳 float 嚴格子類別 __float__() 方法的支援：這些方法將需要回傳 float 的實例。

    • 回傳 complex 嚴格子類別 __complex__() 方法的支援：這些方法將需要回傳 complex 的實例。

    • 將 int() 委派給 __trunc__() 方法。

    • 在 complex() 建構子中將複數作為 real 或 imag 引數傳遞現在已被棄用；它應該只作為單個位置引數傳遞。 (由 Serhiy Storchaka 於 gh-109218 貢獻。)

  • calendar：calendar.January 和 calendar.February 常數已被棄用並被 calendar.JANUARY 和 calendar.FEBRUARY 取代。 (由 Prince Roshan 於 gh-103636 貢獻。)

  • codeobject.co_lnotab：請改用 codeobject.co_lines() 方法。

  • datetime：

    • utcnow()：請改用 datetime.datetime.now(tz=datetime.UTC)。

    • utcfromtimestamp()：請改用 datetime.datetime.fromtimestamp(timestamp, tz=datetime.UTC)。

  • gettext：複數值必須是整數。

  • importlib：

    • load_module() method：請改用 exec_module()。

    • cache_from_source() debug_override 參數已被棄用：請改用 optimization 參數。

  • importlib.metadata：

    • EntryPoints 元組介面。

    • 回傳值上的隱式 None。

  • logging：自 Python 3.3 起，warn() 方法已被棄用，請改用 warning()。

  • mailbox：已棄用 StringIO 輸入和文本模式，請改用 BytesIO 和二進位模式。

  • os：在多執行緒行程中呼叫 os.register_at_fork()。

  • pydoc.ErrorDuringImport：exc_info 參數的元組值已被棄用，請用例外實例。

  • re：現在對正規表示式中的數值群組參照和群組名稱用了更嚴格的規則。現在只有 ASCII 數碼序列被接受作為數值參照。位元組模式和替換字串中的群組名稱現在只能包含 ASCII 字母、數碼和底線。（由 Serhiy Storchaka 於 gh-91760 貢獻。）

  • sre_compile、sre_constants 和 sre_parse 模組。

  • shutil：rmtree() 的 onerror 參數在 Python 3.12 中已被棄用；請改用 onexc 參數。

  • ssl 選項和協定：

    • 不帶協定引數的 ssl.SSLContext 已被棄用。

    • ssl.SSLContext：set_npn_protocols() 和 selected_npn_protocol() 已被棄用：請改用 ALPN。

    • ssl.OP_NO_SSL* 選項

    • ssl.OP_NO_TLS* 選項

    • ssl.PROTOCOL_SSLv3

    • ssl.PROTOCOL_TLS

    • ssl.PROTOCOL_TLSv1

    • ssl.PROTOCOL_TLSv1_1

    • ssl.PROTOCOL_TLSv1_2

    • ssl.TLSVersion.SSLv3

    • ssl.TLSVersion.TLSv1

    • ssl.TLSVersion.TLSv1_1

  • sysconfig.is_python_build() 的 check_home 參數已被棄用並被忽略。

  • threading 方法：

  • typing.Text (gh-92332)。

  • unittest.IsolatedAsyncioTestCase：從測試案例中回傳非 None 的值已被棄用。

  • urllib.parse 已棄用函式：請改用 urlparse()。

    • splitattr()

    • splithost()

    • splitnport()

    • splitpasswd()

    • splitport()

    • splitquery()

    • splittag()

    • splittype()

    • splituser()

    • splitvalue()

    • to_bytes()

  • wsgiref：SimpleHandler.stdout.write() 不應該進行部分寫入。

  • xml.etree.ElementTree：已棄用對 Element 的真值測試。在未來版本中，它將始終回傳 True。請改用明確的 len(elem) 或 elem is not None 測試。

  • zipimport.zipimporter.load_module() 已被棄用：請改用 exec_module()。

已移除

asynchat 和 asyncore

  • These two modules have been removed according to the schedule in PEP 594, having been deprecated in Python 3.6. Use asyncio instead. (Contributed by Nikita Sobolev in gh-96580.)

configparser

distutils

  • Remove the distutils package. It was deprecated in Python 3.10 by PEP 632 "Deprecate distutils module". For projects still using distutils and cannot be updated to something else, the setuptools project can be installed: it still provides distutils. (Contributed by Victor Stinner in gh-92584.)

ensurepip

  • Remove the bundled setuptools wheel from ensurepip, and stop installing setuptools in environments created by venv.

    pip (>= 22.1) does not require setuptools to be installed in the environment. setuptools-based (and distutils-based) packages can still be used with pip install, since pip will provide setuptools in the build environment it uses for building a package.

    easy_install, pkg_resources, setuptools and distutils are no longer provided by default in environments created with venv or bootstrapped with ensurepip, since they are part of the setuptools package. For projects relying on these at runtime, the setuptools project should be declared as a dependency and installed separately (typically, using pip).

    （由 Pradyun Gedam 於 gh-95299 中貢獻。）

enum

  • Remove enum's EnumMeta.__getattr__, which is no longer needed for enum attribute access. (Contributed by Ethan Furman in gh-95083.)

ftplib

  • 移除 ftplib 的 FTP_TLS.ssl_version 類別屬性：請改用 context 參數。（由 Victor Stinner 於 gh-94172 中貢獻。）

gzip

  • Remove the filename attribute of gzip's gzip.GzipFile, deprecated since Python 2.6, use the name attribute instead. In write mode, the filename attribute added '.gz' file extension if it was not present. (Contributed by Victor Stinner in gh-94196.)

hashlib

  • Remove the pure Python implementation of hashlib's hashlib.pbkdf2_hmac(), deprecated in Python 3.10. Python 3.10 and newer requires OpenSSL 1.1.1 (PEP 644): this OpenSSL version provides a C implementation of pbkdf2_hmac() which is faster. (Contributed by Victor Stinner in gh-94199.)

importlib

  • 現已完成清理 importlib 中許多過去已經棄用的東西：

    • 對 module_repr() 的參照和支援已刪除。（由 Barry Warsaw 在 gh-97850 中貢獻。）

    • importlib.util.set_package, importlib.util.set_loader and importlib.util.module_for_loader have all been removed. (Contributed by Brett Cannon and Nikita Sobolev in gh-65961 and gh-97850.)

    • Support for find_loader() and find_module() APIs have been removed. (Contributed by Barry Warsaw in gh-98040.)

    • importlib.abc.Finder, pkgutil.ImpImporter, and pkgutil.ImpLoader have been removed. (Contributed by Barry Warsaw in gh-98040.)

imp

  • The imp module has been removed. (Contributed by Barry Warsaw in gh-98040.)

    To migrate, consult the following correspondence table:

    imp

    importlib

    imp.NullImporter

    將 None 插入 sys.path_importer_cache

    imp.cache_from_source()

    importlib.util.cache_from_source()

    imp.find_module()

    importlib.util.find_spec()

    imp.get_magic()

    importlib.util.MAGIC_NUMBER

    imp.get_suffixes()

    importlib.machinery.SOURCE_SUFFIXES, importlib.machinery.EXTENSION_SUFFIXES, and importlib.machinery.BYTECODE_SUFFIXES

    imp.get_tag()

    sys.implementation.cache_tag

    imp.load_module()

    importlib.import_module()

    imp.new_module(name)

    types.ModuleType(name)

    imp.reload()

    importlib.reload()

    imp.source_from_cache()

    importlib.util.source_from_cache()

    imp.load_source()

    見下文

    用以下取代 imp.load_source()：

    import importlib.util
import importlib.machinery

def load_source(modname, filename):
    loader = importlib.machinery.SourceFileLoader(modname, filename)
    spec = importlib.util.spec_from_file_location(modname, filename, loader=loader)
    module = importlib.util.module_from_spec(spec)
    # The module is always executed and not cached in sys.modules.
    # Uncomment the following line to cache the module.
    # sys.modules[module.__name__] = module
    loader.exec_module(module)
    return module

  • 移除 imp 函式和屬性、沒有替代方案：

    • 未以文件記錄的函式：

      • imp.init_builtin()

      • imp.load_compiled()

      • imp.load_dynamic()

      • imp.load_package()

    • imp.lock_held(), imp.acquire_lock(), imp.release_lock(): the locking scheme has changed in Python 3.3 to per-module locks.

    • imp.find_module() 常數：SEARCH_ERROR、PY_SOURCE、PY_COMPILED、C_EXTENSION、PY_RESOURCE、PKG_DIRECTORY、C_BUILTIN、PY_FROZEN、PY_CODERESOURCE、IMP_HOOK。

io

  • Remove io's io.OpenWrapper and _pyio.OpenWrapper, deprecated in Python 3.10: just use open() instead. The open() (io.open()) function is a built-in function. Since Python 3.10, _pyio.open() is also a static method. (Contributed by Victor Stinner in gh-94169.)

locale

smtpd

  • The smtpd module has been removed according to the schedule in PEP 594, having been deprecated in Python 3.4.7 and 3.5.4. Use the aiosmtpd PyPI module or any other asyncio-based server instead. (Contributed by Oleg Iarygin in gh-93243.)

sqlite3

  • The following undocumented sqlite3 features, deprecated in Python 3.10, are now removed:

    • sqlite3.enable_shared_cache()

    • sqlite3.OptimizedUnicode

    If a shared cache must be used, open the database in URI mode using the cache=shared query parameter.

    The sqlite3.OptimizedUnicode text factory has been an alias for str since Python 3.3. Code that previously set the text factory to OptimizedUnicode can either use str explicitly, or rely on the default value which is also str.

    （由 Erlend E. Aasland 於 gh-92548 中貢獻。）

ssl

  • Remove ssl's ssl.RAND_pseudo_bytes() function, deprecated in Python 3.6: use os.urandom() or ssl.RAND_bytes() instead. (Contributed by Victor Stinner in gh-94199.)

  • Remove the ssl.match_hostname() function. It was deprecated in Python 3.7. OpenSSL performs hostname matching since Python 3.7, Python no longer uses the ssl.match_hostname() function. (Contributed by Victor Stinner in gh-94199.)

  • Remove the ssl.wrap_socket() function, deprecated in Python 3.7: instead, create a ssl.SSLContext object and call its ssl.SSLContext.wrap_socket method. Any package that still uses ssl.wrap_socket() is broken and insecure. The function neither sends a SNI TLS extension nor validates the server hostname. Code is subject to CWE 295 (Improper Certificate Validation). (Contributed by Victor Stinner in gh-94199.)

unittest

webbrowser

  • Remove support for obsolete browsers from webbrowser. The removed browsers include: Grail, Mosaic, Netscape, Galeon, Skipstone, Iceape, Firebird, and Firefox versions 35 and below (gh-102871).

xml.etree.ElementTree

  • Remove the ElementTree.Element.copy() method of the pure Python implementation, deprecated in Python 3.10, use the copy.copy() function instead. The C implementation of xml.etree.ElementTree has no copy() method, only a __copy__() method. (Contributed by Victor Stinner in gh-94383.)

zipimport

  • Remove zipimport's find_loader() and find_module() methods, deprecated in Python 3.10: use the find_spec() method instead. See PEP 451 for the rationale. (Contributed by Victor Stinner in gh-94379.)

其他

  • Remove the suspicious rule from the documentation Makefile and Doc/tools/rstlint.py, both in favor of sphinx-lint. (Contributed by Julien Palard in gh-98179.)

  • Remove the keyfile and certfile parameters from the ftplib, imaplib, poplib and smtplib modules, and the key_file, cert_file and check_hostname parameters from the http.client module, all deprecated since Python 3.6. Use the context parameter (ssl_context in imaplib) instead. (Contributed by Victor Stinner in gh-94172.)

  • 移除數個標準函式庫模組與測試中的 Jython 相容性修補程式。（由 Nikita Sobolev 於 gh-99482 中貢獻。）

  • 移除 ctypes 模組中的 _use_broken_old_ctypes_structure_semantics_ 旗標。（由 Nikita Sobolev 於 gh-99285 中貢獻。）

移植至 Python 3.12

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

Python API 的變更

  • More strict rules are now applied for numerical group references and group names in regular expressions. Only sequence of ASCII digits is now accepted as a numerical reference. The group name in bytes patterns and replacement strings can now only contain ASCII letters and digits and underscore. (Contributed by Serhiy Storchaka in gh-91760.)

  • Remove randrange() functionality deprecated since Python 3.10. Formerly, randrange(10.0) losslessly converted to randrange(10). Now, it raises a TypeError. Also, the exception raised for non-integer values such as randrange(10.5) or randrange('10') has been changed from ValueError to TypeError. This also prevents bugs where randrange(1e25) would silently select from a larger range than randrange(10**25). (Originally suggested by Serhiy Storchaka gh-86388.)

  • argparse.ArgumentParser changed encoding and error handler for reading arguments from file (e.g. fromfile_prefix_chars option) from default text encoding (e.g. locale.getpreferredencoding(False)) to filesystem encoding and error handler. Argument files should be encoded in UTF-8 instead of ANSI Codepage on Windows.

  • Remove the asyncore-based smtpd module deprecated in Python 3.4.7 and 3.5.4. A recommended replacement is the asyncio-based aiosmtpd PyPI module.

  • shlex.split(): Passing None for s argument now raises an exception, rather than reading sys.stdin. The feature was deprecated in Python 3.9. (Contributed by Victor Stinner in gh-94352.)

  • The os module no longer accepts bytes-like paths, like bytearray and memoryview types: only the exact bytes type is accepted for bytes strings. (Contributed by Victor Stinner in gh-98393.)

  • syslog.openlog() and syslog.closelog() now fail if used in subinterpreters. syslog.syslog() may still be used in subinterpreters, but now only if syslog.openlog() has already been called in the main interpreter. These new restrictions do not apply to the main interpreter, so only a very small set of users might be affected. This change helps with interpreter isolation. Furthermore, syslog is a wrapper around process-global resources, which are best managed from the main interpreter. (Contributed by Donghee Na in gh-99127.)

  • The undocumented locking behavior of cached_property() is removed, because it locked across all instances of the class, leading to high lock contention. This means that a cached property getter function could now run more than once for a single instance, if two threads race. For most simple cached properties (e.g. those that are idempotent and simply calculate a value based on other attributes of the instance) this will be fine. If synchronization is needed, implement locking within the cached property getter function or around multi-threaded access points.

  • sys._current_exceptions() now returns a mapping from thread-id to an exception instance, rather than to a (typ, exc, tb) tuple. (Contributed by Irit Katriel in gh-103176.)

  • When extracting tar files using tarfile or shutil.unpack_archive(), pass the filter argument to limit features that may be surprising or dangerous. See Extraction filters for details.

  • The output of the tokenize.tokenize() and tokenize.generate_tokens() functions is now changed due to the changes introduced in PEP 701. This means that STRING tokens are not emitted any more for f-strings and the tokens described in PEP 701 are now produced instead: FSTRING_START, FSTRING_MIDDLE and FSTRING_END are now emitted for f-string "string" parts in addition to the appropriate tokens for the tokenization in the expression components. For example for the f-string f"start {1+1} end" the old version of the tokenizer emitted:

    1,0-1,18:           STRING         'f"start {1+1} end"'

    while the new version emits:

    1,0-1,2:            FSTRING_START  'f"'
1,2-1,8:            FSTRING_MIDDLE 'start '
1,8-1,9:            OP             '{'
1,9-1,10:           NUMBER         '1'
1,10-1,11:          OP             '+'
1,11-1,12:          NUMBER         '1'
1,12-1,13:          OP             '}'
1,13-1,17:          FSTRING_MIDDLE ' end'
1,17-1,18:          FSTRING_END    '"'

    Additionally, there may be some minor behavioral changes as a consequence of the changes required to support PEP 701. Some of these changes include:

    • The type attribute of the tokens emitted when tokenizing some invalid Python characters such as ! has changed from ERRORTOKEN to OP.

    • Incomplete single-line strings now also raise tokenize.TokenError as incomplete multiline strings do.

    • Some incomplete or invalid Python code now raises tokenize.TokenError instead of returning arbitrary ERRORTOKEN tokens when tokenizing it.

    • Mixing tabs and spaces as indentation in the same file is not supported anymore and will raise a TabError.

  • The threading module now expects the _thread module to have an _is_main_interpreter attribute. It is a function with no arguments that returns True if the current interpreter is the main interpreter.

    Any library or application that provides a custom _thread module should provide _is_main_interpreter(). (See gh-112826.)

建置變更

  • Python no longer uses setup.py to build shared C extension modules. Build parameters like headers and libraries are detected in configure script. Extensions are built by Makefile. Most extensions use pkg-config and fall back to manual detection. (Contributed by Christian Heimes in gh-93939.)

  • va_start() with two parameters, like va_start(args, format), is now required to build Python. va_start() is no longer called with a single parameter. (Contributed by Kumar Aditya in gh-93207.)

  • CPython now uses the ThinLTO option as the default link time optimization policy if the Clang compiler accepts the flag. (Contributed by Donghee Na in gh-89536.)

  • Add COMPILEALL_OPTS variable in Makefile to override compileall options (default: -j0) in make install. Also merged the 3 compileall commands into a single command to build .pyc files for all optimization levels (0, 1, 2) at once. (Contributed by Victor Stinner in gh-99289.)

  • Add platform triplets for 64-bit LoongArch:

    • loongarch64-linux-gnusf

    • loongarch64-linux-gnuf32

    • loongarch64-linux-gnu

    （由 Zhang Na 於 gh-90656 中貢獻。）

  • PYTHON_FOR_REGEN now require Python 3.10 or newer.

  • Autoconf 2.71 and aclocal 1.16.4 is now required to regenerate !configure. (Contributed by Christian Heimes in gh-89886.)

  • Windows builds and macOS installers from python.org now use OpenSSL 3.0.

C API 變更

新增功能

  • PEP 683: Introduce Immortal Objects, which allows objects to bypass reference counts, and related changes to the C-API:

    • _Py_IMMORTAL_REFCNT：定義物件的參照計數

      as immortal.

    • _Py_IsImmortal Checks if an object has the immortal reference count.

    • PyObject_HEAD_INIT 這現在將初始化參照計數

      _Py_IMMORTAL_REFCNT（與 Py_BUILD_CORE 一起使用時）。

    • SSTATE_INTERNED_IMMORTAL An identifier for interned unicode objects

      that are immortal.

    • SSTATE_INTERNED_IMMORTAL_STATIC An identifier for interned unicode

      objects that are immortal and static

    • sys.getunicodeinternedsize 這會回傳 unicode 的總數

      objects that have been interned. This is now needed for refleak.py to correctly track reference counts and allocated blocks

    （由 Eddie Elizondo 於 gh-84436 中貢獻。）

  • PEP 684: Add the new Py_NewInterpreterFromConfig() function and PyInterpreterConfig, which may be used to create sub-interpreters with their own GILs. (See PEP 684: A Per-Interpreter GIL for more info.) (Contributed by Eric Snow in gh-104110.)

  • In the limited C API version 3.12, Py_INCREF() and Py_DECREF() functions are now implemented as opaque function calls to hide implementation details. (Contributed by Victor Stinner in gh-105387.)

移植至 Python 3.12

  • Legacy Unicode APIs based on Py_UNICODE* representation has been removed. Please migrate to APIs based on UTF-8 or wchar_t*.

  • Argument parsing functions like PyArg_ParseTuple() doesn't support Py_UNICODE* based format (e.g. u, Z) anymore. Please migrate to other formats for Unicode like s, z, es, and U.

  • tp_weaklist for all static builtin types is always NULL. This is an internal-only field on PyTypeObject but we're pointing out the change in case someone happens to be accessing the field directly anyway. To avoid breakage, consider using the existing public C-API instead, or, if necessary, the (internal-only) _PyObject_GET_WEAKREFS_LISTPTR() macro.

  • This internal-only PyTypeObject.tp_subclasses may now not be a valid object pointer. Its type was changed to void* to reflect this. We mention this in case someone happens to be accessing the internal-only field directly.

    To get a list of subclasses, call the Python method __subclasses__() (using PyObject_CallMethod(), for example).

  • Add support of more formatting options (left aligning, octals, uppercase hexadecimals, intmax_t, ptrdiff_t, wchar_t C strings, variable width and precision) in PyUnicode_FromFormat() and PyUnicode_FromFormatV(). (Contributed by Serhiy Storchaka in gh-98836.)

  • An unrecognized format character in PyUnicode_FromFormat() and PyUnicode_FromFormatV() now sets a SystemError. In previous versions it caused all the rest of the format string to be copied as-is to the result string, and any extra arguments discarded. (Contributed by Serhiy Storchaka in gh-95781.)

  • Fix wrong sign placement in PyUnicode_FromFormat() and PyUnicode_FromFormatV(). (Contributed by Philip Georgi in gh-95504.)

  • Extension classes wanting to add a __dict__ or weak reference slot should use Py_TPFLAGS_MANAGED_DICT and Py_TPFLAGS_MANAGED_WEAKREF instead of tp_dictoffset and tp_weaklistoffset, respectively. The use of tp_dictoffset and tp_weaklistoffset is still supported, but does not fully support multiple inheritance (gh-95589), and performance may be worse. Classes declaring Py_TPFLAGS_MANAGED_DICT must call _PyObject_VisitManagedDict() and _PyObject_ClearManagedDict() to traverse and clear their instance's dictionaries. To clear weakrefs, call PyObject_ClearWeakRefs(), as before.

  • The PyUnicode_FSDecoder() function no longer accepts bytes-like paths, like bytearray and memoryview types: only the exact bytes type is accepted for bytes strings. (Contributed by Victor Stinner in gh-98393.)

  • The Py_CLEAR, Py_SETREF and Py_XSETREF macros now only evaluate their arguments once. If an argument has side effects, these side effects are no longer duplicated. (Contributed by Victor Stinner in gh-98724.)

  • The interpreter's error indicator is now always normalized. This means that PyErr_SetObject(), PyErr_SetString() and the other functions that set the error indicator now normalize the exception before storing it. (Contributed by Mark Shannon in gh-101578.)

  • _Py_RefTotal is no longer authoritative and only kept around for ABI compatibility. Note that it is an internal global and only available on debug builds. If you happen to be using it then you'll need to start using _Py_GetGlobalRefTotal().

  • The following functions now select an appropriate metaclass for the newly created type:

    Creating classes whose metaclass overrides tp_new is deprecated, and in Python 3.14+ it will be disallowed. Note that these functions ignore tp_new of the metaclass, possibly allowing incomplete initialization.

    Note that PyType_FromMetaclass() (added in Python 3.12) already disallows creating classes whose metaclass overrides tp_new (__new__() in Python).

    Since tp_new overrides almost everything PyType_From* functions do, the two are incompatible with each other. The existing behavior -- ignoring the metaclass for several steps of type creation -- is unsafe in general, since (meta)classes assume that tp_new was called. There is no simple general workaround. One of the following may work for you:

    • If you control the metaclass, avoid using tp_new in it:

      • If initialization can be skipped, it can be done in tp_init instead.

      • If the metaclass doesn't need to be instantiated from Python, set its tp_new to NULL using the Py_TPFLAGS_DISALLOW_INSTANTIATION flag. This makes it acceptable for PyType_From* functions.

    • Avoid PyType_From* functions: if you don't need C-specific features (slots or setting the instance size), create types by calling the metaclass.

    • If you know the tp_new can be skipped safely, filter the deprecation warning out using warnings.catch_warnings() from Python.

  • PyOS_InputHook and PyOS_ReadlineFunctionPointer are no longer called in subinterpreters. This is because clients generally rely on process-wide global state (since these callbacks have no way of recovering extension module state).

    This also avoids situations where extensions may find themselves running in a subinterpreter that they don't support (or haven't yet been loaded in). See gh-104668 for more info.

  • PyLongObject has had its internals changed for better performance. Although the internals of PyLongObject are private, they are used by some extension modules. The internal fields should no longer be accessed directly, instead the API functions beginning PyLong_... should be used instead. Two new unstable API functions are provided for efficient access to the value of PyLongObjects which fit into a single machine word:

  • Custom allocators, set via PyMem_SetAllocator(), are now required to be thread-safe, regardless of memory domain. Allocators that don't have their own state, including "hooks", are not affected. If your custom allocator is not already thread-safe and you need guidance then please create a new GitHub issue and CC @ericsnowcurrently.

已棄用

Pending removal in Python 3.14

Pending removal in Python 3.15

Pending removal in future versions

下列 API 已被棄用並將會被移除，不過目前尚未訂定移除日期。

已移除

  • Remove the token.h header file. There was never any public tokenizer C API. The token.h header file was only designed to be used by Python internals. (Contributed by Victor Stinner in gh-92651.)

  • Legacy Unicode APIs have been removed. See PEP 623 for detail.

    • PyUnicode_WCHAR_KIND

    • PyUnicode_AS_UNICODE()

    • PyUnicode_AsUnicode()

    • PyUnicode_AsUnicodeAndSize()

    • PyUnicode_AS_DATA()

    • PyUnicode_FromUnicode()

    • PyUnicode_GET_SIZE()

    • PyUnicode_GetSize()

    • PyUnicode_GET_DATA_SIZE()

  • 移除 PyUnicode_InternImmortal() 函式巨集。（由 Victor Stinner 於 gh-85858 中貢獻。）