ABC causing massive memory swell

まず、提供された再現コードを影響を受ける Python バージョンで実行し、issubclass ループ中のメモリを測定します。関係する ABCMeta と issubclass の動作を読み、その後、結果を issue #92810 および pydantic #3829 と比較します。キャッシュの増大とサブクラスツリーの走査に対処し、ABC の登録動作を後退させない、合意済みでテスト済みのアプローチが得られれば完了です。

Bug report

from abc import ABC, ABCMeta
from datetime import datetime


abcclasses = set()

normalclasses = set()

for i in range(10000):
    abcclasses.add(ABCMeta("abc_"+str(i), (ABC, ), {}))
    normalclasses.add(type("normal_"+str(i), (object,), {}))


if __name__ == '__main__':

    starttime = datetime.now()
    import os, psutil
    process = psutil.Process(os.getpid())
    mb = 1024 * 1024
    mem = last = process.memory_info().rss
    print(f'{i + 1:>4d} {mem / mb:8.2f}MB {(mem - last) / mb:+8.2f}MB | {"━" * int(mem / 8_000_000)}')
    for item in normalclasses:
        issubclass(item, ABC)

    mem = process.memory_info().rss
    print(f'{i + 1:>4d} {mem / mb:8.2f}MB {(mem - last) / mb:+8.2f}MB | {"━" * int(mem / 8_000_000)}')
    print(f"This took {datetime.now()-starttime}")

Your environment

Running this takes several minutes and consumes over 20 gbs of memory.

I've checked it against python 3.8-3.10. The issue is just that ABC caches every single issubclass evaluation combined with it searching its entire subclass tree recursively causes it to grind to a halt if you have any significant number of subclass 10k sub classes of ABC and 10k normal classes being compared to the base ABC climbs to over 20 GBs. I am not really sure it is necessary to check the entire subclass tree. I presume it is because of concerns about register calls happening low in the class tree.

It would likely be far more efficient to chase upward the calls to register and place those in the parent classes then to search the entire class hierarchy. This can get really aggressive performance issues if you have any multiple inheritance structures going on as well as those will get checked multiple times.

See these issues.
https://github.com/python/cpython/issues/92810
https://github.com/samuelcolvin/pydantic/issues/3829