Improve import time of various stdlib modules

[AlexWaygood]

Feature or enhancement

Proposal:

As noted in https://discuss.python.org/t/deferred-computation-evalution-for-toplevels-imports-and-dataclasses/34173, typing isn't the slowest stdlib module in terms of import time, but neither is it one of the quickest. We should speed it up, if possible.

Links to previous discussion of this feature:

https://discuss.python.org/t/deferred-computation-evalution-for-toplevels-imports-and-dataclasses/34173

Linked PRs

• gh-109653: typing.py: improve import time by creating soft-deprecated members on demand #109651
• gh-109653: Improve enum import time by avoiding import of functools #109789
• gh-109653: Remove unused imports in the Lib/ directory #109803
• gh-109653: Avoid a top-level import of types in functools #109804
• gh-109653: Just import recursive_repr in dataclasses #109822
• gh-109653: Improve the import time of email.utils #109824
• GH-109653: Defer import of importlib.metadata._adapters #109829
• gh-109653: Fix py312 regression in the import time of random #110221
• [3.12] gh-109653: Fix py312 regression in the import time of random (GH-110221) #110247
• gh-109653: Defer importing warnings in several modules #110286
• gh-109653: Improve import time of logging by lazy loading traceback #112995
• gh-109653: Speedup import of threading module #114509
• gh-109653: Improve import time of importlib.metadata / email.utils #114664
• gh-109653: Reduce import overhead of uuid module on Linux #115160
• gh-109653: What's new: Note improved import times for several stdlib modules in Python 3.13 #118697

[AlexWaygood]

Retitling this issue to reflect a change of scope: part of the reason for the slow import time of typing is the slow import time of some of the modules it imports. There may be low-hanging fruit in some other stdlib modules as well.

[hugovk]

An idea: review the output of a linter for "unused" imports.

Some may be false positives due to needing an import's side effects, but I expect there will be some that are no longer needed.

[itamaro]

Some may be false positives due to needing an import's side effects ...

How do you feel about making an intentional effort to get rid of such reliance on side effects in the stdlib?
Explicit is better than implicit :)

[AlexWaygood]

An idea: review the output of a linter for "unused" imports.

FWIW, I can only find one in the stdlib that I'm 100% confident with, using ruff/pycln:

cpython/Lib/asyncio/tasks.py

Line 20 in e8be0c9

import warnings

I think this is because Victor went through them all fairly recently, in #105411. There are a fair few in Lib/test/, though. I'll file a PR in a bit.

[sobolevn]

This is also interesting:

• dataclasses copies recursive_repr to avoid a dependency:

  cpython/Lib/dataclasses.py

  Lines 248 to 266 in 51863b7

  	# This function's logic is copied from "recursive_repr" function in
  	# reprlib module to avoid dependency.
  	def _recursive_repr(user_function):
  	# Decorator to make a repr function return "..." for a recursive
  	# call.
  	repr_running = set()
  	@functools.wraps(user_function)
  	def wrapper(self):
  	key = id(self), _thread.get_ident()
  	if key in repr_running:
  	return '...'
  	repr_running.add(key)
  	try:
  	result = user_function(self)
  	finally:
  	repr_running.discard(key)
  	return result
  	return wrapper

• But, it imports functools:

  cpython/Lib/dataclasses.py

  Line 7 in 51863b7

  import functools

• And it imports reprlib.recursive_repr:

  cpython/Lib/functools.py

  Line 20 in 51863b7

  from reprlib import recursive_repr