`ProcessPoolExecutor` fails to notify when called with `shutdown(cancel_futures=True)`

[chrisvanrun]

Bug report

Bug description:

I am running a small processor that handles large tasks. When a task encounters an exception I want to store all errors (including other tasks errors) and cancel any other queued work.

However, it seems like the ProcessPoolExecutor's _ExecutorManagerThread does not notify any waiting threads of the future cancelations: making the idiom as_completed hang indefinitely.

Minimal Reproducable Example

import concurrent.futures
import time


def task(n: int) -> int:
 if n == 2:
 raise Exception("Not gonna do it")
 else:
 time.sleep(0.1)
 return n


def main() -> None:
 with concurrent.futures.ProcessPoolExecutor(
 max_workers=2,
 ) as executor:
 futures = [executor.submit(task, i) for i in range(1, 16)]

 for future in concurrent.futures.as_completed(futures):
 try:
 result = future.result()
 except Exception as e:
 print(f"Exception: {e}")
 executor.shutdown(wait=False, cancel_futures=True)
 continue
 print(f"Result: {result}")

 # Look for 'CANCELLED' here:
 print("Other futures states:", [f._state for f in futures])


if __name__ == "__main__":
 main()

Workaround

A workaround is to break after the first future that completed with a failure and do a post hoc gathering of other exceptions of non-cancelled futures.

Suggested Fix

In

cpython/Lib/concurrent/futures/process.py

Lines 520 to 522 in a68ddea

	for work_id, work_item in self.pending_work_items.items():
	if not work_item.future.cancel():
	new_pending_work_items[work_id] = work_item

One could replace it with:

for work_id, work_item in self.pending_work_items.items():
 canceled = work_item.future.cancel()
 if canceled:
 work_item.future.set_running_or_notify_cancel()
 else:
 new_pending_work_items[work_id] = work_item

Similar like:

• gh-109934: notify cancelled futures on thread pool shutdown #134618

CPython versions tested on:

3.13

Operating systems tested on:

macOS

Linked PRs

• gh-136655: ensure cancelled futures are notified on process pool shutdown #140021

[duaneg]

Yep, looks like the same sort of issue as #109934. The unit test I added for that can't be used directly for process pool, but I think I've managed to come up with one that works reliably based on your repro. It will fail for the thread pool executor without the fix for #109934, so I'll base the PR on top of that branch.

[YvesDup]

Except if I missed something, you should open a PR including your fix and a test.

[duaneg]

Except if I missed something, you should open a PR including your fix and a test.

I added the fix for this issue to #134618, which is the PR I opened to fix #109934. I did it that way rather than open a separate PR as otherwise the test will fail on the thread pool executor due to that bug, and I thought it likely a maintainer/reviewer would prefer to deal with them together.

If it is helpful I'd be happy to open a separate PR just for this issue, though. We can skip the test for the thread pool executor, with a TODO to re-enable it once #109934 is fixed.

[YvesDup]

I suggest opening separately 2 PRs with tests.

[duaneg]

I've not updated the PR for the other issue yet, but would be happy to do so in whatever way is most convenient for reviewers/maintainers.