Worker management race conditions in `ThreadPoolExecutor` and `ProcessPoolExecutor`

We have tests that ThreadPoolExecutor and ProcessPoolExecutor reuse workers. The test wait for the previous task to complete before submitting the next task.

https://github.com/python/cpython/blob/29829b58a8328a7c2ccacaa74c1d7d120a5e5ca5/Lib/test/test_concurrent_futures/test_thread_pool.py#L44-L50

The problem is that the tasks are potentially completed before the thread or process is marked as idle. Calling Future.result() doesn't guarantee that there is an idle worker. This happens much more frequently with the GIL disabled, but it can happen with the GIL enabled too if you have an unlucky GIL switch.

https://github.com/python/cpython/blob/29829b58a8328a7c2ccacaa74c1d7d120a5e5ca5/Lib/concurrent/futures/thread.py#L81-L95

For example, you can reliably trigger the problem with the GIL enabled by adding a short time.sleep(0.01) immediately after the work_item.run() call in:
https://github.com/python/cpython/blob/29829b58a8328a7c2ccacaa74c1d7d120a5e5ca5/Lib/concurrent/futures/thread.py#L92

It's not clear to me if this is just a problem with the unit tests or an actual bug with the executor implementation.

Linked PRs

• gh-117355

Commencez par les tests de réutilisation cités dans Lib/test/test_concurrent_futures/test_thread_pool.py et par le code du cycle de vie des workers dans Lib/concurrent/futures/thread.py ; examinez également les chemins correspondants de ProcessPoolExecutor. Déterminez si la condition de concurrence nécessite une synchronisation plus stricte des tests ou une modification de l’executor, puis exécutez la suite de tests concurrent-futures et assurez-vous que la réutilisation des workers est fiable aussi bien lors d’une exécution avec le GIL activé que lors d’une exécution avec le GIL désactivé.