What’s New In Python 3.8

Editor

Raymond Hettinger

This article explains the new features in Python 3.8, compared to 3.7. For full details, see the changelog.

Prerelease users should be aware that this document is currently in draft form. It will be updated as Python 3.8 moves towards release, so it’s worth checking back even after reading earlier versions. Some notable items not yet covered are:

  • PEP 578 - Runtime audit hooks for potentially sensitive operations

  • python -m asyncio runs a natively async REPL

Summary – Release highlights

New Features

Assignment expressions

There is new syntax := that assigns values to variables as part of a larger expression. It is affectionately known as “walrus operator” due to its resemblance to the eyes and tusks of a walrus.

In this example, the assignment expression helps avoid calling len() twice:

if (n := len(a)) > 10:
    print(f"List is too long ({n} elements, expected <= 10)")

A similar benefit arises during regular expression matching where match objects are needed twice, once to test whether a match occurred and another to extract a subgroup:

discount = 0.0
if (mo := re.search(r'(\d+)% discount', advertisement)):
    discount = float(mo.group(1)) / 100.0

The operator is also useful with while-loops that compute a value to test loop termination and then need that same value again in the body of the loop:

# Loop over fixed length blocks
while (block := f.read(256)) != '':
    process(block)

Another motivating use case arises in list comprehensions where a value computed in a filtering condition is also needed in the expression body:

[clean_name.title() for name in names
 if (clean_name := normalize('NFC', name)) in allowed_names]

Try to limit use of the walrus operator to clean cases that reduce complexity and improve readability.

See PEP 572 for a full description.

(Contributed by Emily Morehouse in bpo-35224.)

Positional-only parameters

There is a new function parameter syntax / to indicate that some function parameters must be specified positionally and cannot be used as keyword arguments. This is the same notation shown by help() for C functions annotated with Larry Hastings’ Argument Clinic tool.

In the following example, parameters a and b are positional-only, while c or d can be positional or keyword, and e or f are required to be keywords:

def f(a, b, /, c, d, *, e, f):
    print(a, b, c, d, e, f)

The following is a valid call:

f(10, 20, 30, d=40, e=50, f=60)

However, these are invalid calls:

f(10, b=20, c=30, d=40, e=50, f=60)   # b cannot be a keyword argument
f(10, 20, 30, 40, 50, f=60)           # e must be a keyword argument

One use case for this notation is that it allows pure Python functions to fully emulate behaviors of existing C coded functions. For example, the built-in pow() function does not accept keyword arguments:

def pow(x, y, z=None, /):
    "Emulate the built in pow() function"
    r = x ** y
    return r if z is None else r%z

Another use case is to preclude keyword arguments when the parameter name is not helpful. For example, the builtin len() function has the signature len(obj, /). This precludes awkward calls such as:

len(obj='hello')  # The "obj" keyword argument impairs readability

A further benefit of marking a parameter as positional-only is that it allows the parameter name to be changed in the future without risk of breaking client code. For example, in the statistics module, the parameter name dist may be changed in the future. This was made possible with the following function specification:

def quantiles(dist, /, *, n=4, method='exclusive')
    ...

Since the parameters to the left of / are not exposed as possible keywords, the parameters names remain available for use in **kwargs:

>>> def f(a, b, /, **kwargs):
...     print(a, b, kwargs)
...
>>> f(10, 20, a=1, b=2, c=3)         # a and b are used in two ways
10 20 {'a': 1, 'b': 2, 'c': 3}

This greatly simplifies the implementation of functions and methods that need to accept arbitrary keyword arguments. For example, here is an except from code in the collections module:

class Counter(dict):

    def __init__(self, iterable=None, /, **kwds):
        # Note "iterable" is a possible keyword argument

See PEP 570 for a full description.

(Contributed by Pablo Galindo in bpo-36540.)

Parallel filesystem cache for compiled bytecode files

The new PYTHONPYCACHEPREFIX setting (also available as -X pycache_prefix) configures the implicit bytecode cache to use a separate parallel filesystem tree, rather than the default __pycache__ subdirectories within each source directory.

The location of the cache is reported in sys.pycache_prefix (None indicates the default location in __pycache__ subdirectories).

(Contributed by Carl Meyer in bpo-33499.)

Debug build uses the same ABI as release build

Python now uses the same ABI whether it built in release or debug mode. On Unix, when Python is built in debug mode, it is now possible to load C extensions built in release mode and C extensions built using the stable ABI.

Release builds and debug builds are now ABI compatible: defining the Py_DEBUG macro no longer implies the Py_TRACE_REFS macro, which introduces the only ABI incompatibility. The Py_TRACE_REFS macro, which adds the sys.getobjects() function and the PYTHONDUMPREFS environment variable, can be set using the new ./configure --with-trace-refs build option. (Contributed by Victor Stinner in bpo-36465.)

On Unix, C extensions are no longer linked to libpython except on Android and Cygwin. It is now possible for a statically linked Python to load a C extension built using a shared library Python. (Contributed by Victor Stinner in bpo-21536.)

On Unix, when Python is built in debug mode, import now also looks for C extensions compiled in release mode and for C extensions compiled with the stable ABI. (Contributed by Victor Stinner in bpo-36722.)

To embed Python into an application, a new --embed option must be passed to python3-config --libs --embed to get -lpython3.8 (link the application to libpython). To support both 3.8 and older, try python3-config --libs --embed first and fallback to python3-config --libs (without --embed) if the previous command fails.

Add a pkg-config python-3.8-embed module to embed Python into an application: pkg-config python-3.8-embed --libs includes -lpython3.8. To support both 3.8 and older, try pkg-config python-X.Y-embed --libs first and fallback to pkg-config python-X.Y --libs (without --embed) if the previous command fails (replace X.Y with the Python version).

On the other hand, pkg-config python3.8 --libs no longer contains -lpython3.8. C extensions must not be linked to libpython (except on Android and Cygwin, whose cases are handled by the script); this change is backward incompatible on purpose. (Contributed by Victor Stinner in bpo-36721.)

f-strings support = for self-documenting expressions and debugging

Added an = specifier to f-strings. An f-string such as f'{expr=}' will expand to the text of the expression, an equal sign, then the representation of the evaluated expression. For example:

>>> user = 'eric_idle'
>>> member_since = date(1975, 7, 31)
>>> f'{user=} {member_since=}'
"user='eric_idle' member_since=datetime.date(1975, 7, 31)"

The usual f-string format specifiers allow more control over how the result of the expression is displayed:

>>> delta = date.today() - member_since
>>> f'{user=!s}  {delta.days=:,d}'
'user=eric_idle  delta.days=16,075'

The = specifier will display the whole expression so that calculations can be shown:

>>> print(f'{theta=}  {cos(radians(theta))=:.3f}')
theta=30  cos(radians(theta))=0.866

(Contributed by Eric V. Smith and Larry Hastings in bpo-36817.)

PEP 587: Python Initialization Configuration

The PEP 587 adds a new C API to configure the Python Initialization providing finer control on the whole configuration and better error reporting.

New structures:

New functions:

This PEP also adds _PyRuntimeState.preconfig (PyPreConfig type) and PyInterpreterState.config (PyConfig type) fields to these internal structures. PyInterpreterState.config becomes the new reference configuration, replacing global configuration variables and other private variables.

See Python Initialization Configuration for the documentation.

See PEP 587 for a full description.

(Contributed by Victor Stinner in bpo-36763.)

Vectorcall: a fast calling protocol for CPython

The “vectorcall” protocol is added to the Python/C API. It is meant to formalize existing optimizations which were already done for various classes. Any extension type implementing a callable can use this protocol.

This is currently provisional, the aim is to make it fully public in Python 3.9.

See PEP 590 for a full description.

(Contributed by Jeroen Demeyer and Mark Shannon in bpo-36974.)

Pickle protocol 5 with out-of-band data buffers

When pickle is used to transfer large data between Python processes in order to take advantage of multi-core or multi-machine processing, it is important to optimize the transfer by reducing memory copies, and possibly by applying custom techniques such as data-dependent compression.

The pickle protocol 5 introduces support for out-of-band buffers where PEP 3118-compatible data can be transmitted separately from the main pickle stream, at the discretion of the communication layer.

See PEP 574 for a full description.

(Contributed by Antoine Pitrou in bpo-36785.)

Other Language Changes

  • A continue statement was illegal in the finally clause due to a problem with the implementation. In Python 3.8 this restriction was lifted. (Contributed by Serhiy Storchaka in bpo-32489.)

  • The bool, int, and fractions.Fraction types now have an as_integer_ratio() method like that found in float and decimal.Decimal. This minor API extension makes it possible to write numerator, denominator = x.as_integer_ratio() and have it work across multiple numeric types. (Contributed by Lisa Roach in bpo-33073 and Raymond Hettinger in bpo-37819.)

  • Constructors of int, float and complex will now use the __index__() special method, if available and the corresponding method __int__(), __float__() or __complex__() is not available. (Contributed by Serhiy Storchaka in bpo-20092.)

  • Added support of \N{name} escapes in regular expressions:

    >>> notice = 'Copyright © 2019'
>>> copyright_year_pattern = re.compile(r'\N{copyright sign}\s*(\d{4})')
>>> int(copyright_year_pattern.search(notice).group(1))
2019

    (Contributed by Jonathan Eunice and Serhiy Storchaka in bpo-30688.)

  • Dict and dictviews are now iterable in reversed insertion order using reversed(). (Contributed by Rémi Lapeyre in bpo-33462.)

  • The syntax allowed for keyword names in function calls was further restricted. In particular, f((keyword)=arg) is no longer allowed. It was never intended to permit more than a bare name on the left-hand side of a keyword argument assignment term. See bpo-34641.

  • Generalized iterable unpacking in yield and return statements no longer requires enclosing parentheses. This brings the yield and return syntax into better agreement with normal assignment syntax:

    >>> def parse(family):
        lastname, *members = family.split()
        return lastname.upper(), *members

>>> parse('simpsons homer marge bart lisa sally')
('SIMPSONS', 'homer', 'marge', 'bart', 'lisa', 'sally')

    (Contributed by David Cuthbert and Jordan Chapman in bpo-32117.)

  • When a comma is missed in code such as [(10, 20) (30, 40)], the compiler displays a SyntaxWarning with a helpful suggestion. This improves on just having a TypeError indicating that the first tuple was not callable. (Contributed by Serhiy Storchaka in bpo-15248.)

  • Arithmetic operations between subclasses of datetime.date or datetime.datetime and datetime.timedelta objects now return an instance of the subclass, rather than the base class. This also affects the return type of operations whose implementation (directly or indirectly) uses datetime.timedelta arithmetic, such as datetime.datetime.astimezone(). (Contributed by Paul Ganssle in bpo-32417.)

  • When the Python interpreter is interrupted by Ctrl-C (SIGINT) and the resulting KeyboardInterrupt exception is not caught, the Python process now exits via a SIGINT signal or with the correct exit code such that the calling process can detect that it died due to a Ctrl-C. Shells on POSIX and Windows use this to properly terminate scripts in interactive sessions. (Contributed by Google via Gregory P. Smith in bpo-1054041.)

  • Some advanced styles of programming require updating the types.CodeType object for an existing function. Since code objects are immutable, a new code object needs to be created, one that is modeled on the existing code object. With 19 parameters, this was somewhat tedious. Now, the new replace() method makes it possible to create a clone with a few altered parameters.

    Here’s an example that alters the statistics.mean() function to prevent the data parameter from being used as a keyword argument:

    >>> from statistics import mean
>>> mean(data=[10, 20, 90])
40
>>> mean.__code__ = mean.__code__.replace(co_posonlyargcount=1)
>>> mean(data=[10, 20, 90])
Traceback (most recent call last):
  ...
TypeError: mean() got some positional-only arguments passed as keyword arguments: 'data'

    (Contributed by Victor Stinner in bpo-37032.)

  • For integers, the three-argument form of the pow() function now permits the exponent to be negative in the case where the base is relatively prime to the modulus. It then computes a modular inverse to the base when the exponent is -1, and a suitable power of that inverse for other negative exponents. For example, to compute the modular multiplicative inverse of 38 modulo 137, write:

    >>> pow(38, -1, 137)
119
>>> 119 * 38 % 137
1

    Modular inverses arise in the solution of linear Diophantine equations. For example, to find integer solutions for 4258𝑥 + 147𝑦 = 369, first rewrite as 4258𝑥 ≡ 369 (mod 147) then solve:

    >>> x = 369 * pow(4258, -1, 147) % 147
>>> y = (4258 * x - 369) // -147
>>> 4258 * x + 147 * y
369

    (Contributed by Mark Dickinson in bpo-36027.)

  • Dict comprehensions have been synced-up with dict literals so that the key is computed first and the value second:

    >>> # Dict comprehension
>>> cast = {input('role? '): input('actor? ') for i in range(2)}
role? King Arthur
actor? Chapman
role? Black Knight
actor? Cleese

>>> # Dict literal
>>> cast = {input('role? '): input('actor? ')}
role? Sir Robin
actor? Eric Idle

    The guaranteed execution order is helpful with assignment expressions because variables assigned in the key expression will be available in the value expression:

    >>> names = ['Martin von Löwis', 'Łukasz Langa', 'Walter Dörwald']
>>> {(n := normalize('NFC', name)).casefold() : n for name in names}
{'martin von löwis': 'Martin von Löwis',
 'łukasz langa': 'Łukasz Langa',
 'walter dörwald': 'Walter Dörwald'}

New Modules

  • The new importlib.metadata module provides (provisional) support for reading metadata from third-party packages. For example, it can extract an installed package’s version number, list of entry points, and more:

    >>> # Note following example requires that the popular "requests"
>>> # package has been installed.
>>>
>>> from importlib.metadata import version, requires, files
>>> version('requests')
'2.22.0'
>>> list(requires('requests'))
['chardet (<3.1.0,>=3.0.2)']
>>> list(files('requests'))[:5]
[PackagePath('requests-2.22.0.dist-info/INSTALLER'),
 PackagePath('requests-2.22.0.dist-info/LICENSE'),
 PackagePath('requests-2.22.0.dist-info/METADATA'),
 PackagePath('requests-2.22.0.dist-info/RECORD'),
 PackagePath('requests-2.22.0.dist-info/WHEEL')]

    (Contributed in bpo-34632 by Barry Warsaw and Jason R. Coombs.)

Improved Modules

ast

AST nodes now have end_lineno and end_col_offset attributes, which give the precise location of the end of the node. (This only applies to nodes that have lineno and col_offset attributes.)

The ast.parse() function has some new flags:

  • type_comments=True causes it to return the text of PEP 484 and PEP 526 type comments associated with certain AST nodes;

  • mode='func_type' can be used to parse PEP 484 “signature type comments” (returned for function definition AST nodes);

  • feature_version=(3, N) allows specifying an earlier Python 3 version. (For example, feature_version=(3, 4) will treat async and await as non-reserved words.)

New function ast.get_source_segment() returns the source code for a specific AST node.

asyncio

On Windows, the default event loop is now ProactorEventLoop. (Contributed by Victor Stinner in bpo-34687.)

ProactorEventLoop now also supports UDP. (Contributed by Adam Meily and Andrew Svetlov in bpo-29883.)

ProactorEventLoop can now be interrupted by KeyboardInterrupt (“CTRL+C”). (Contributed by Vladimir Matveev in bpo-23057.)

builtins

The compile() built-in has been improved to accept the ast.PyCF_ALLOW_TOP_LEVEL_AWAIT flag. With this new flag passed, compile() will allow top-level await, async for and async with constructs that are usually considered invalid syntax. Asynchronous code object marked with the CO_COROUTINE flag may then be returned.

(Contributed by Matthias Bussonnier in bpo-34616)

collections

The _asdict() method for collections.namedtuple() now returns a dict instead of a collections.OrderedDict. This works because regular dicts have guaranteed ordering since Python 3.7. If the extra features of OrderedDict are required, the suggested remediation is to cast the result to the desired type: OrderedDict(nt._asdict()). (Contributed by Raymond Hettinger in bpo-35864.)

ctypes

On Windows, CDLL and subclasses now accept a winmode parameter to specify flags for the underlying LoadLibraryEx call. The default flags are set to only load DLL dependencies from trusted locations, including the path where the DLL is stored (if a full or partial path is used to load the initial DLL) and paths added by add_dll_directory().

functools

functools.lru_cache() can now be used as a straight decorator rather than as a function returning a decorator. So both of these are now supported:

@lru_cache
def f(x):
    ...

@lru_cache(maxsize=256)
def f(x):
    ...

(Contributed by Raymond Hettinger in bpo-36772.)

datetime

Added new alternate constructors datetime.date.fromisocalendar() and datetime.datetime.fromisocalendar(), which construct date and datetime objects respectively from ISO year, week number and weekday; these are the inverse of each class’s isocalendar method. (Contributed by Paul Ganssle in bpo-36004.)

gettext

Added pgettext() and its variants. (Contributed by Franz Glasner, Éric Araujo, and Cheryl Sabella in bpo-2504.)

idlelib and IDLE

Output over N lines (50 by default) is squeezed down to a button. N can be changed in the PyShell section of the General page of the Settings dialog. Fewer, but possibly extra long, lines can be squeezed by right clicking on the output. Squeezed output can be expanded in place by double-clicking the button or into the clipboard or a separate window by right-clicking the button. (Contributed by Tal Einat in bpo-1529353.)

Add “Run Customized” to the Run menu to run a module with customized settings. Any command line arguments entered are added to sys.argv. They also re-appear in the box for the next customized run. One can also suppress the normal Shell main module restart. (Contributed by Cheryl Sabella, Terry Jan Reedy, and others in bpo-5680 and bpo-37627.)

Add optional line numbers for IDLE editor windows. Windows open without line numbers unless set otherwise in the General tab of the configuration dialog. Line numbers for an existing window are shown and hidden in the Options menu. (Contributed by Tal Einat and Saimadhav Heblikar in bpo-17535.)

The changes above have been backported to 3.7 maintenance releases.

inspect

The inspect.getdoc() function can now find docstrings for __slots__ if that attribute is a dict where the values are docstrings. This provides documentation options similar to what we already have for property(), classmethod(), and staticmethod():

class AudioClip:
    __slots__ = {'bit_rate': 'expressed in kilohertz to one decimal place',
                 'duration': 'in seconds, rounded up to an integer'}
    def __init__(self, bit_rate, duration):
        self.bit_rate = round(bit_rate / 1000.0, 1)
        self.duration = ceil(duration)

io

In development mode (-X env) and in debug build, the io.IOBase finalizer now logs the exception if the close() method fails. The exception is ignored silently by default in release build. (Contributed by Victor Stinner in bpo-18748.)

gc

get_objects() can now receive an optional generation parameter indicating a generation to get objects from. Contributed in bpo-36016 by Pablo Galindo.

gzip

Added the mtime parameter to gzip.compress() for reproducible output. (Contributed by Guo Ci Teo in bpo-34898.)

A BadGzipFile exception is now raised instead of OSError for certain types of invalid or corrupt gzip files. (Contributed by Filip Gruszczyński, Michele Orrù, and Zackery Spytz in bpo-6584.)

idlelib and IDLE

Add optional line numbers for IDLE editor windows. Windows open without line numbers unless set otherwise in the General tab of the configuration dialog. (Contributed by Tal Einat and Saimadhav Heblikar in bpo-17535.)

Output over N lines (50 by default) is squeezed down to a button. N can be changed in the PyShell section of the General page of the Settings dialog. Fewer, but possibly extra long, lines can be squeezed by right clicking on the output. Squeezed output can be expanded in place by double-clicking the button or into the clipboard or a separate window by right-clicking the button. (Contributed by Tal Einat in bpo-1529353.)

The changes above have been backported to 3.7 maintenance releases.

json.tool

Add option --json-lines to parse every input line as separate JSON object. (Contributed by Weipeng Hong in bpo-31553.)

math

Added new function math.dist() for computing Euclidean distance between two points. (Contributed by Raymond Hettinger in bpo-33089.)

Expanded the math.hypot() function to handle multiple dimensions. Formerly, it only supported the 2-D case. (Contributed by Raymond Hettinger in bpo-33089.)

Added new function, math.prod(), as analogous function to sum() that returns the product of a ‘start’ value (default: 1) times an iterable of numbers:

>>> prior = 0.8
>>> likelihoods = [0.625, 0.84, 0.30]
>>> (link: http://math.prod) math.prod(likelihoods, start=prior)
0.126

(Contributed by Pablo Galindo in bpo-35606)

Added new function math.isqrt() for computing integer square roots. (Contributed by Mark Dickinson in bpo-36887.)

The function math.factorial() no longer accepts arguments that are not int-like. (Contributed by Pablo Galindo in bpo-33083.)

mmap

The mmap.mmap class now has an madvise() method to access the madvise() system call. (Contributed by Zackery Spytz in bpo-32941.)

multiprocessing

Added new multiprocessing.shared_memory module. (Contributed Davin Potts in bpo-35813.)

On macOS, the spawn start method is now used by default. (Contributed by Victor Stinner in bpo-33725.)

os

Added new function add_dll_directory() on Windows for providing additional search paths for native dependencies when importing extension modules or loading DLLs using ctypes.

A new os.memfd_create() function was added to wrap the memfd_create() syscall. (Contributed by Zackery Spytz and Christian Heimes in bpo-26836.)

os.path

os.path functions that return a boolean result like exists(), lexists(), isdir(), isfile(), islink(), and ismount() now return False instead of raising ValueError or its subclasses UnicodeEncodeError and UnicodeDecodeError for paths that contain characters or bytes unrepresentable at the OS level. (Contributed by Serhiy Storchaka in bpo-33721.)

expanduser() on Windows now prefers the USERPROFILE environment variable and does not use HOME, which is not normally set for regular user accounts.

ncurses

Added a new variable holding structured version information for the underlying ncurses library: ncurses_version. (Contributed by Serhiy Storchaka in bpo-31680.)

pathlib

pathlib.Path methods that return a boolean result like exists(), is_dir(), is_file(), is_mount(), is_symlink(), is_block_device(), is_char_device(), is_fifo(), is_socket() now return False instead of raising ValueError or its subclass UnicodeEncodeError for paths that contain characters unrepresentable at the OS level. (Contributed by Serhiy Storchaka in bpo-33721.)

Added pathlib.Path.link_to() which creates a hard link pointing to a path. (Contributed by Joannah Nanjekye in bpo-26978)

pickle

Reduction methods can now include a 6th item in the tuple they return. This item should specify a custom state-setting method that’s called instead of the regular __setstate__ method. (Contributed by Pierre Glaser and Olivier Grisel in bpo-35900)

pickle extensions subclassing the C-optimized Pickler can now override the pickling logic of functions and classes by defining the special reducer_override() method. (Contributed by Pierre Glaser and Olivier Grisel in bpo-35900)

plistlib

Added new plistlib.UID and enabled support for reading and writing NSKeyedArchiver-encoded binary plists. (Contributed by Jon Janzen in bpo-26707.)

py_compile

py_compile.compile() now supports silent mode. (Contributed by Joannah Nanjekye in bpo-22640.)

socket

Added create_server() and has_dualstack_ipv6() convenience functions to automate the necessary tasks usually involved when creating a server socket, including accepting both IPv4 and IPv6 connections on the same socket. (Contributed by Giampaolo Rodola in bpo-17561.)

The socket.if_nameindex(), socket.if_nametoindex(), and socket.if_indextoname() functions have been implemented on Windows. (Contributed by Zackery Spytz in bpo-37007.)

shlex

The new shlex.join() function acts as the inverse of shlex.split(). (Contributed by Bo Bayles in bpo-32102.)

shutil

shutil.copytree() now accepts a new dirs_exist_ok keyword argument. (Contributed by Josh Bronson in bpo-20849.)

shutil.make_archive() now defaults to the modern pax (POSIX.1-2001) format for new archives to improve portability and standards conformance, inherited from the corresponding change to the tarfile module. (Contributed by C.A.M. Gerlach in bpo-30661.)

ssl

Added SSLContext.post_handshake_auth to enable and ssl.SSLSocket.verify_client_post_handshake() to initiate TLS 1.3 post-handshake authentication. (Contributed by Christian Heimes in bpo-34670.)

statistics

Added statistics.fmean() as a faster, floating point variant of statistics.mean(). (Contributed by Raymond Hettinger and Steven D’Aprano in bpo-35904.)

Added statistics.geometric_mean() (Contributed by Raymond Hettinger in bpo-27181.)

Added statistics.multimode() that returns a list of the most common values. (Contributed by Raymond Hettinger in bpo-35892.)

Added statistics.quantiles() that divides data or a distribution in to equiprobable intervals (e.g. quartiles, deciles, or percentiles). (Contributed by Raymond Hettinger in bpo-36546.)

Added statistics.NormalDist, a tool for creating and manipulating normal distributions of a random variable. (Contributed by Raymond Hettinger in bpo-36018.)

>>> temperature_feb = NormalDist.from_samples([4, 12, -3, 2, 7, 14])
>>> temperature_feb.mean
6.0
>>> temperature_feb.stdev
6.356099432828281

>>> temperature_feb.cdf(3)            # Chance of being under 3 degrees
0.3184678262814532
>>> # Relative chance of being 7 degrees versus 10 degrees
>>> temperature_feb.pdf(7) / temperature_feb.pdf(10)
1.2039930378537762

>>> el_niño = NormalDist(4, 2.5)
>>> temperature_feb += el_niño        # Add in a climate effect
>>> temperature_feb
NormalDist(mu=10.0, sigma=6.830080526611674)

>>> temperature_feb * (9/5) + 32      # Convert to Fahrenheit
NormalDist(mu=50.0, sigma=12.294144947901014)
>>> temperature_feb.samples(3)        # Generate random samples
[7.672102882379219, 12.000027119750287, 4.647488369766392]

sys

Add new sys.unraisablehook() function which can be overridden to control how “unraisable exceptions” are handled. It is called when an exception has occurred but there is no way for Python to handle it. For example, when a destructor raises an exception or during garbage collection (gc.collect()). (Contributed by Victor Stinner in bpo-36829.)

tarfile

The tarfile module now defaults to the modern pax (POSIX.1-2001) format for new archives, instead of the previous GNU-specific one. This improves cross-platform portability with a consistent encoding (UTF-8) in a standardized and extensible format, and offers several other benefits. (Contributed by C.A.M. Gerlach in bpo-36268.)

threading

tokenize

The tokenize module now implicitly emits a NEWLINE token when provided with input that does not have a trailing new line. This behavior now matches what the C tokenizer does internally. (Contributed by Ammar Askar in bpo-33899.)

tkinter

Added methods selection_from(), selection_present(), selection_range() and selection_to() in the tkinter.Spinbox class. (Contributed by Juliette Monsel in bpo-34829.)

Added method moveto() in the tkinter.Canvas class. (Contributed by Juliette Monsel in bpo-23831.)

The tkinter.PhotoImage class now has transparency_get() and transparency_set() methods. (Contributed by Zackery Spytz in bpo-25451.)

time

Added new clock CLOCK_UPTIME_RAW for macOS 10.12. (Contributed by Joannah Nanjekye in bpo-35702.)

typing

The typing module incorporates several new features:

unicodedata

unittest

  • Added AsyncMock to support an asynchronous version of Mock. Appropriate new assert functions for testing have been added as well. (Contributed by Lisa Roach in bpo-26467).

  • Added addModuleCleanup() and addClassCleanup() to unittest to support cleanups for setUpModule() and setUpClass(). (Contributed by Lisa Roach in bpo-24412.)

  • Several mock assert functions now also print a list of actual calls upon failure. (Contributed by Petter Strandmark in bpo-35047.)

venv

  • venv now includes an Activate.ps1 script on all platforms for activating virtual environments under PowerShell Core 6.1. (Contributed by Brett Cannon in bpo-32718.)

weakref

  • The proxy objects returned by weakref.proxy() now support the matrix multiplication operators @ and @= in addition to the other numeric operators. (Contributed by Mark Dickinson in bpo-36669.)

xml

  • As mitigation against DTD and external entity retrieval, the xml.dom.minidom and xml.sax modules no longer process external entities by default. (Contributed by Christian Heimes in bpo-17239.)

  • The .find*() methods in the xml.etree.ElementTree module support wildcard searches like {*}tag which ignores the namespace and {namespace}* which returns all tags in the given namespace. (Contributed by Stefan Behnel in bpo-28238.)

  • The xml.etree.ElementTree module provides a new function –xml.etree.ElementTree.canonicalize() that implements C14N 2.0. (Contributed by Stefan Behnel in bpo-13611.)

  • The target object of xml.etree.ElementTree.XMLParser can receive namespace declaration events through the new callback methods start_ns() and end_ns(). Additionally, the xml.etree.ElementTree.TreeBuilder target can be configured to process events about comments and processing instructions to include them in the generated tree. (Contributed by Stefan Behnel in bpo-36676 and bpo-36673.)

Optimizations

  • The subprocess module can now use the os.posix_spawn() function in some cases for better performance. Currently, it is only used on macOS and Linux (using glibc 2.24 or newer) if all these conditions are met:

    • close_fds is false;

    • preexec_fn, pass_fds, cwd and start_new_session parameters are not set;

    • the executable path contains a directory.

    (Contributed by Joannah Nanjekye and Victor Stinner in bpo-35537.)

  • shutil.copyfile(), shutil.copy(), shutil.copy2(), shutil.copytree() and shutil.move() use platform-specific “fast-copy” syscalls on Linux and macOS in order to copy the file more efficiently. “fast-copy” means that the copying operation occurs within the kernel, avoiding the use of userspace buffers in Python as in “outfd.write(infd.read())”. On Windows shutil.copyfile() uses a bigger default buffer size (1 MiB instead of 16 KiB) and a memoryview()-based variant of shutil.copyfileobj() is used. The speedup for copying a 512 MiB file within the same partition is about +26% on Linux, +50% on macOS and +40% on Windows. Also, much less CPU cycles are consumed. See Platform-dependent efficient copy operations section. (Contributed by Giampaolo Rodola’ in bpo-33671.)

  • shutil.copytree() uses os.scandir() function and all copy functions depending from it use cached os.stat() values. The speedup for copying a directory with 8000 files is around +9% on Linux, +20% on Windows and +30% on a Windows SMB share. Also the number of os.stat() syscalls is reduced by 38% making shutil.copytree() especially faster on network filesystems. (Contributed by Giampaolo Rodola’ in bpo-33695.)

  • The default protocol in the pickle module is now Protocol 4, first introduced in Python 3.4. It offers better performance and smaller size compared to Protocol 3 available since Python 3.0.

  • Removed one Py_ssize_t member from PyGC_Head. All GC tracked objects (e.g. tuple, list, dict) size is reduced 4 or 8 bytes. (Contributed by Inada Naoki in bpo-33597)

  • uuid.UUID now uses __slots__ to reduce its memory footprint.

  • Improved performance of operator.itemgetter() by 33%. Optimized argument handling and added a fast path for the common case of a single non-negative integer index into a tuple (which is the typical use case in the standard library). (Contributed by Raymond Hettinger in bpo-35664.)

  • Sped-up field lookups in collections.namedtuple(). They are now more than two times faster, making them the fastest form of instance variable lookup in Python. (Contributed by Raymond Hettinger, Pablo Galindo, and Joe Jevnik, Serhiy Storchaka in bpo-32492.)

  • The list constructor does not overallocate the internal item buffer if the input iterable has a known length (the input implements __len__). This makes the created list 12% smaller on average. (Contributed by Raymond Hettinger and Pablo Galindo in bpo-33234.)

  • Doubled the speed of class variable writes. When a non-dunder attribute was updated, there was an unnecessary call to update slots. (Contributed by Stefan Behnel, Pablo Galindo Salgado, Raymond Hettinger, Neil Schemenauer, and Serhiy Storchaka in bpo-36012.)

  • Reduced an overhead of converting arguments passed to many builtin functions and methods. This sped up calling some simple builtin functions and methods up to 20–50%. (Contributed by Serhiy Storchaka in bpo-23867, bpo-35582 and bpo-36127.)

  • LOAD_GLOBAL instruction now uses new “per opcode cache” mechanism. It is about 40% faster now. (Contributed by Yury Selivanov and Inada Naoki in bpo-26219.)

Build and C API Changes

  • Default sys.abiflags became an empty string: the m flag for pymalloc became useless (builds with and without pymalloc are ABI compatible) and so has been removed. (Contributed by Victor Stinner in bpo-36707.)

    Example of changes:

    • Only python3.8 program is installed, python3.8m program is gone.

    • Only python3.8-config script is installed, python3.8m-config script is gone.

    • The m flag has been removed from the suffix of dynamic library filenames: extension modules in the standard library as well as those produced and installed by third-party packages, like those downloaded from PyPI. On Linux, for example, the Python 3.7 suffix .cpython-37m-x86_64-linux-gnu.so became .cpython-38-x86_64-linux-gnu.so in Python 3.8.

  • The header files have been reorganized to better separate the different kinds of APIs:

    • Include/*.h should be the portable public stable C API.

    • Include/cpython/*.h should be the unstable C API specific to CPython; public API, with some private API prefixed by _Py or _PY.

    • Include/internal/*.h is the private internal C API very specific to CPython. This API comes with no backward compatibility warranty and should not be used outside CPython. It is only exposed for very specific needs like debuggers and profiles which has to access to CPython internals without calling functions. This API is now installed by make install.

    (Contributed by Victor Stinner in bpo-35134 and bpo-35081, work initiated by Eric Snow in Python 3.7)

  • Some macros have been converted to static inline functions: parameter types and return type are well defined, they don’t have issues specific to macros, variables have a local scopes. Examples:

    (Contributed by Victor Stinner in bpo-35059.)

  • The PyByteArray_Init() and PyByteArray_Fini() functions have been removed. They did nothing since Python 2.7.4 and Python 3.2.0, were excluded from the limited API (stable ABI), and were not documented. (Contributed by Victor Stinner in bpo-35713.)

  • The result of PyExceptionClass_Name() is now of type const char * rather of char *. (Contributed by Serhiy Storchaka in bpo-33818.)

  • The duality of Modules/Setup.dist and Modules/Setup has been removed. Previously, when updating the CPython source tree, one had to manually copy Modules/Setup.dist (inside the source tree) to Modules/Setup (inside the build tree) in order to reflect any changes upstream. This was of a small benefit to packagers at the expense of a frequent annoyance to developers following CPython development, as forgetting to copy the file could produce build failures.

    Now the build system always reads from Modules/Setup inside the source tree. People who want to customize that file are encouraged to maintain their changes in a git fork of CPython or as patch files, as they would do for any other change to the source tree.

    (Contributed by Antoine Pitrou in bpo-32430.)

  • Functions that convert Python number to C integer like PyLong_AsLong() and argument parsing functions like PyArg_ParseTuple() with integer converting format units like 'i' will now use the __index__() special method instead of __int__(), if available. The deprecation warning will be emitted for objects with the __int__() method but without the __index__() method (like Decimal and Fraction). PyNumber_Check() will now return 1 for objects implementing __index__(). PyNumber_Long(), PyNumber_Float() and PyFloat_AsDouble() also now use the __index__() method if available. (Contributed by Serhiy Storchaka in bpo-36048 and bpo-20092.)

  • Heap-allocated type objects will now increase their reference count in PyObject_Init() (and its parallel macro PyObject_INIT) instead of in PyType_GenericAlloc(). Types that modify instance allocation or deallocation may need to be adjusted. (Contributed by Eddie Elizondo in bpo-35810.)

  • The new function PyCode_NewWithPosOnlyArgs() allows to create code objects like PyCode_New(), but with an extra posonlyargcount parameter for indicating the number of positional-only arguments. (Contributed by Pablo Galindo in bpo-37221.)

Deprecated

API and Feature Removals

The following features and APIs have been removed from Python 3.8:

  • The macpath module, deprecated in Python 3.7, has been removed. (Contributed by Victor Stinner in bpo-35471.)

  • The function platform.popen() has been removed, it was deprecated since Python 3.3: use os.popen() instead. (Contributed by Victor Stinner in bpo-35345.)

  • The function time.clock() has been removed, it was deprecated since Python 3.3: use time.perf_counter() or time.process_time() instead, depending on your requirements, to have a well defined behavior. (Contributed by Matthias Bussonnier in bpo-36895.)

  • The pyvenv script has been removed in favor of python3.8 -m venv to help eliminate confusion as to what Python interpreter the pyvenv script is tied to. (Contributed by Brett Cannon in bpo-25427.)

  • parse_qs, parse_qsl, and escape are removed from cgi module. They are deprecated from Python 3.2 or older.

  • filemode function is removed from tarfile module. It is not documented and deprecated since Python 3.3.

  • The XMLParser constructor no longer accepts the html argument. It never had effect and was deprecated in Python 3.4. All other parameters are now keyword-only. (Contributed by Serhiy Storchaka in bpo-29209.)

  • Removed the doctype() method of XMLParser. (Contributed by Serhiy Storchaka in bpo-29209.)

  • “unicode_internal” codec is removed. (Contributed by Inada Naoki in bpo-36297.)

  • The Cache and Statement objects of the sqlite3 module are not exposed to the user. (Contributed by Aviv Palivoda in bpo-30262.)

  • The bufsize keyword argument of fileinput.input() and fileinput.FileInput() which was ignored and deprecated since Python 3.6 has been removed. bpo-36952 (Contributed by Matthias Bussonnier)

  • The functions sys.set_coroutine_wrapper() and sys.get_coroutine_wrapper() deprecated in Python 3.7 have been removed; bpo-36933 (Contributed by Matthias Bussonnier)

Porting to Python 3.8

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

Changes in Python behavior

  • Yield expressions (both yield and yield from clauses) are now disallowed in comprehensions and generator expressions (aside from the iterable expression in the leftmost for clause). (Contributed by Serhiy Storchaka in bpo-10544.)

  • The compiler now produces a SyntaxWarning when identity checks (is and is not) are used with certain types of literals (e.g. strings, ints). These can often work by accident in CPython, but are not guaranteed by the language spec. The warning advises users to use equality tests (== and !=) instead. (Contributed by Serhiy Storchaka in bpo-34850.)

  • The CPython interpreter can swallow exceptions in some circumstances. In Python 3.8 this happens in less cases. In particular, exceptions raised when getting the attribute from the type dictionary are no longer ignored. (Contributed by Serhiy Storchaka in bpo-35459.)

  • Removed __str__ implementations from builtin types bool, int, float, complex and few classes from the standard library. They now inherit __str__() from object. As result, defining the __repr__() method in the subclass of these classes will affect they string representation. (Contributed by Serhiy Storchaka in bpo-36793.)

  • On AIX, sys.platform doesn’t contain the major version anymore. It is always 'aix', instead of 'aix3' .. 'aix7'. Since older Python versions include the version number, it is recommended to always use the sys.platform.startswith('aix'). (Contributed by M. Felt in bpo-36588.)

  • PyEval_AcquireLock() and PyEval_AcquireThread() now terminate the current thread if called while the interpreter is finalizing, making them consistent with PyEval_RestoreThread(), Py_END_ALLOW_THREADS(), and PyGILState_Ensure(). If this behaviour is not desired, guard the call by checking _Py_IsFinalizing() or sys.is_finalizing().

Changes in the Python API

  • The os.getcwdb() function now uses the UTF-8 encoding on Windows, rather than the ANSI code page: see PEP 529 for the rationale. The function is no longer deprecated on Windows. (Contributed by Victor Stinner in bpo-37412.)

  • subprocess.Popen can now use os.posix_spawn() in some cases for better performance. On Windows Subsystem for Linux and QEMU User Emulation, Popen constructor using os.posix_spawn() no longer raise an exception on errors like missing program, but the child process fails with a non-zero returncode. (Contributed by Joannah Nanjekye and Victor Stinner in bpo-35537.)

  • The imap.IMAP4.logout() method no longer ignores silently arbitrary exceptions.

  • The function platform.popen() has been removed, it was deprecated since Python 3.3: use os.popen() instead. (Contributed by Victor Stinner in bpo-35345.)

  • The statistics.mode() function no longer raises an exception when given multimodal data. Instead, it returns the first mode encountered in the input data. (Contributed by Raymond Hettinger in bpo-35892.)

  • The selection() method of the tkinter.ttk.Treeview class no longer takes arguments. Using it with arguments for changing the selection was deprecated in Python 3.6. Use specialized methods like selection_set() for changing the selection. (Contributed by Serhiy Storchaka in bpo-31508.)

  • The writexml(), toxml() and toprettyxml() methods of the xml.dom.minidom module, and xml.etree now preserve the attribute order specified by the user. (Contributed by Diego Rojas and Raymond Hettinger in bpo-34160.)

  • A dbm.dumb database opened with flags 'r' is now read-only. dbm.dumb.open() with flags 'r' and 'w' no longer creates a database if it does not exist. (Contributed by Serhiy Storchaka in bpo-32749.)

  • The doctype() method defined in a subclass of XMLParser will no longer be called and will cause emitting a RuntimeWarning instead of a DeprecationWarning. Define the doctype() method on a target for handling an XML doctype declaration. (Contributed by Serhiy Storchaka in bpo-29209.)

  • A RuntimeError is now raised when the custom metaclass doesn’t provide the __classcell__ entry in the namespace passed to type.__new__. A DeprecationWarning was emitted in Python 3.6–3.7. (Contributed by Serhiy Storchaka in bpo-23722.)

  • The cProfile.Profile class can now be used as a context manager. (Contributed by Scott Sanderson in bpo-29235.)

  • shutil.copyfile(), shutil.copy(), shutil.copy2(), shutil.copytree() and shutil.move() use platform-specific “fast-copy” syscalls (see Platform-dependent efficient copy operations section).

  • shutil.copyfile() default buffer size on Windows was changed from 16 KiB to 1 MiB.

  • The PyGC_Head struct has changed completely. All code that touched the struct member should be rewritten. (See bpo-33597)

  • The PyInterpreterState struct has been moved into the “internal” header files (specifically Include/internal/pycore_pystate.h). An opaque PyInterpreterState is still available as part of the public API (and stable ABI). The docs indicate that none of the struct’s fields are public, so we hope no one has been using them. However, if you do rely on one or more of those private fields and have no alternative then please open a BPO issue. We’ll work on helping you adjust (possibly including adding accessor functions to the public API). (See bpo-35886.)

  • Asyncio tasks can now be named, either by passing the name keyword argument to asyncio.create_task() or the create_task() event loop method, or by calling the set_name() method on the task object. The task name is visible in the repr() output of asyncio.Task and can also be retrieved using the get_name() method.

  • The mmap.flush() method now returns None on success and raises an exception on error under all platforms. Previously, its behavior was platform-depended: a nonzero value was returned on success; zero was returned on error under Windows. A zero value was returned on success; an exception was raised on error under Unix. (Contributed by Berker Peksag in bpo-2122.)

  • xml.dom.minidom and xml.sax modules no longer process external entities by default. (Contributed by Christian Heimes in bpo-17239.)

  • Deleting a key from a read-only dbm database (dbm.dumb, dbm.gnu or dbm.ndbm) raises error (dbm.dumb.error, dbm.gnu.error or dbm.ndbm.error) instead of KeyError. (Contributed by Xiang Zhang in bpo-33106.)

  • expanduser() on Windows now prefers the USERPROFILE environment variable and does not use HOME, which is not normally set for regular user accounts.

  • DLL dependencies for extension modules and DLLs loaded with ctypes on Windows are now resolved more securely. Only the system paths, the directory containing the DLL or PYD file, and directories added with add_dll_directory() are searched for load-time dependencies. Specifically, PATH and the current working directory are no longer used, and modifications to these will no longer have any effect on normal DLL resolution. If your application relies on these mechanisms, you should check for add_dll_directory() and if it exists, use it to add your DLLs directory while loading your library. Note that Windows 7 users will need to ensure that Windows Update KB2533625 has been installed (this is also verified by the installer). (See bpo-36085.)

  • The header files and functions related to pgen have been removed after its replacement by a pure Python implementation. (Contributed by Pablo Galindo in bpo-36623.)

  • types.CodeType has a new parameter in the second position of the constructor (posonlyargcount) to support positional-only arguments defined in PEP 570. The first argument (argcount) now represents the total number of positional arguments (including positional-only arguments). A new replace() method of types.CodeType can be used to make the code future-proof.

Changes in the C API

  • The PyCompilerFlags structure gets a new cf_feature_version field. It should be initialized to PY_MINOR_VERSION. The field is ignored by default, it is used if and only if PyCF_ONLY_AST flag is set in cf_flags.

  • The PyEval_ReInitThreads() function has been removed from the C API. It should not be called explicitly: use PyOS_AfterFork_Child() instead. (Contributed by Victor Stinner in bpo-36728.)

  • On Unix, C extensions are no longer linked to libpython except on Android and Cygwin. When Python is embedded, libpython must not be loaded with RTLD_LOCAL, but RTLD_GLOBAL instead. Previously, using RTLD_LOCAL, it was already not possible to load C extensions which were not linked to libpython, like C extensions of the standard library built by the *shared* section of Modules/Setup. (Contributed by Victor Stinner in bpo-21536.)

  • Use of # variants of formats in parsing or building value (e.g. PyArg_ParseTuple(), Py_BuildValue(), PyObject_CallFunction(), etc.) without PY_SSIZE_T_CLEAN defined raises DeprecationWarning now. It will be removed in 3.10 or 4.0. Read Parsing arguments and building values for detail. (Contributed by Inada Naoki in bpo-36381.)

  • Instances of heap-allocated types (such as those created with PyType_FromSpec()) hold a reference to their type object. Increasing the reference count of these type objects has been moved from PyType_GenericAlloc() to the more low-level functions, PyObject_Init() and PyObject_INIT(). This makes types created through PyType_FromSpec() behave like other classes in managed code.

    Statically allocated types are not affected.

    For the vast majority of cases, there should be no side effect. However, types that manually increase the reference count after allocating an instance (perhaps to work around the bug) may now become immortal. To avoid this, these classes need to call Py_DECREF on the type object during instance deallocation.

    To correctly port these types into 3.8, please apply the following changes:

    • Remove Py_INCREF on the type object after allocating an instance - if any. This may happen after calling PyObject_New(), PyObject_NewVar(), PyObject_GC_New(), PyObject_GC_NewVar(), or any other custom allocator that uses PyObject_Init() or PyObject_INIT().

      Example:

      static foo_struct *
foo_new(PyObject *type) {
    foo_struct *foo = PyObject_GC_New(foo_struct, (PyTypeObject *) type);
    if (foo == NULL)
        return NULL;
#if PY_VERSION_HEX < 0x03080000
    // Workaround for Python issue 35810; no longer necessary in Python 3.8
    PY_INCREF(type)
#endif
    return foo;
}

    • Ensure that all custom tp_dealloc functions of heap-allocated types decrease the type’s reference count.

      Example:

      static void
foo_dealloc(foo_struct *instance) {
    PyObject *type = Py_TYPE(instance);
    PyObject_GC_Del(instance);
#if PY_VERSION_HEX >= 0x03080000
    // This was not needed before Python 3.8 (Python issue 35810)
    Py_DECREF(type);
#endif
}

    (Contributed by Eddie Elizondo in bpo-35810.)

  • The Py_DEPRECATED() macro has been implemented for MSVC. The macro now must be placed before the symbol name.

    Example:

    Py_DEPRECATED(3.8) PyAPI_FUNC(int) Py_OldFunction(void);

    (Contributed by Zackery Spytz in bpo-33407.)

  • The interpreter does not pretend to support binary compatibility of extension types across feature releases, anymore. A PyTypeObject exported by a third-party extension module is supposed to have all the slots expected in the current Python version, including tp_finalize (Py_TPFLAGS_HAVE_FINALIZE is not checked anymore before reading tp_finalize).

    (Contributed by Antoine Pitrou in bpo-32388.)

  • The PyCode_New() has a new parameter in the second position (posonlyargcount) to support PEP 570, indicating the number of positional-only arguments.

  • The functions PyNode_AddChild() and PyParser_AddToken() now accept two additional int arguments end_lineno and end_col_offset.

  • The libpython38.a file to allow MinGW tools to link directly against python38.dll is no longer included in the regular Windows distribution. If you require this file, it may be generated with the gendef and dlltool tools, which are part of the MinGW binutils package:

    gendef python38.dll > tmp.def
dlltool --dllname python38.dll --def tmp.def --output-lib libpython38.a

    The location of an installed pythonXY.dll will depend on the installation options and the version and language of Windows. See Using Python on Windows for more information. The resulting library should be placed in the same directory as pythonXY.lib, which is generally the libs directory under your Python installation.

CPython bytecode changes

  • The interpreter loop has been simplified by moving the logic of unrolling the stack of blocks into the compiler. The compiler emits now explicit instructions for adjusting the stack of values and calling the cleaning-up code for break, continue and return.

    Removed opcodes BREAK_LOOP, CONTINUE_LOOP, SETUP_LOOP and SETUP_EXCEPT. Added new opcodes ROT_FOUR, BEGIN_FINALLY, CALL_FINALLY and POP_FINALLY. Changed the behavior of END_FINALLY and WITH_CLEANUP_START.

    (Contributed by Mark Shannon, Antoine Pitrou and Serhiy Storchaka in bpo-17611.)

  • Added new opcode END_ASYNC_FOR for handling exceptions raised when awaiting a next item in an async for loop. (Contributed by Serhiy Storchaka in bpo-33041.)

  • The MAP_ADD now expects the value as the first element in the stack and the key as the second element. This change was made so the key is always evaluated before the value in dictionary comprehensions, as proposed by PEP 572. (Contributed by Jörn Heissler in bpo-35224.)

Demos and Tools

  • Added a benchmark script for timing various ways to access variables: Tools/scripts/var_access_benchmark.py. (Contributed by Raymond Hettinger in bpo-35884.)