Bug report

Bug description:

This is a self-contained reproducer of a real issue I encountered (see below):

import numbers
from concurrent.futures import ThreadPoolExecutor
from time import time

mylist = [1.0] * 100_000

def check(_):
    for x in mylist:
        isinstance(x, numbers.Integral)

for cores in [1, 2, 4, 8]:
    start = time()
    with ThreadPoolExecutor(cores) as pool:
        list(pool.map(check, range(cores)))
    print(cores, time() - start)

I ran this on my computer, which has 12 cores. I would expect this to the same amount of time regardless of number of threads, since they ought to be run in parallel. In fact, the output looks like this:

Python 3.14t:

1 0.030498981475830078
2 0.07532048225402832
4 0.17853212356567383
8 0.5576419830322266

Python 3.15t (3.15rc2):

1 0.032598018646240234
2 0.05810952186584473
4 0.14814376831054688
8 0.31418490409851074

Original real-world issue

I discovered this issue while benchmarking some scikit-learn code. In particular it's caused by code that uses https://github.com/scikit-learn/scikit-learn/blob/d6f188097e255822d99633f13f4d6304cc76a7c8/sklearn/utils/_missing.py#L40 on all the values in a data structure.

In practice I have optimized away much of the usage of this function, so in future versions of scikit-learn (post-1.9) it hopefully won't be a bottleneck in practice. But it's definitely a real bottleneck in sklearn 1.9, and presumably other people may encounter it in other code.

Potential source of bottleneck

Looking at the profile output of samply suggests _in_weak_set's critical section, maybe (thanks to @ngoldbaum for the link:

CPython versions tested on:

3.14, 3.15

Operating systems tested on:

Linux

Linked PRs

• gh-157526: Call __instancecheck__ and __subclasscheck__ without creating a bound method #157670