gh-80961: Add daemon parameter to ThreadPoolExecutor

[hniksic]

See https://bugs.python.org/issue36780 (GH-80961).

https://bugs.python.org/issue36780

[brianquinlan]

shutdown is defined in Executor because Executor is the abstract superclass for both ThreadPoolExecutor and ProcessPoolExecutor. Unless there is a very strong reason not to, this method should work the same in both executors.

[hniksic]

@brianquinlan This was intentional - I tested shutdown(wait=False) with ProcessPoolExecutor, and found that it raised exceptions and hanged the process at exit. (Not just hanged in the sense of waiting for the pending futures, but completely hanged, even when the futures exited.) So the new functionality is only available in and documented for ThreadPoolExecutor.

For example, when I run this script on Python 3.7:

import time, concurrent.futures

pool = concurrent.futures.ProcessPoolExecutor()

pool.submit(time.sleep, 5)
print(1)

pool.shutdown(wait=False)
print(2)

The expected behavior is for the program to print 1 and 2 and then to wait for 5 seconds before exiting. Instead, it prints 1 and 2, but hangs at exit with the following output:

$ python3.7 ~/Desktop/x
1
2
Error in atexit._run_exitfuncs:
Traceback (most recent call last):
  File "/usr/lib/python3.7/concurrent/futures/process.py", line 101, in _python_exit
    thread_wakeup.wakeup()
  File "/usr/lib/python3.7/concurrent/futures/process.py", line 89, in wakeup
    self._writer.send_bytes(b"")
  File "/usr/lib/python3.7/multiprocessing/connection.py", line 183, in send_bytes
    self._check_closed()
  File "/usr/lib/python3.7/multiprocessing/connection.py", line 136, in _check_closed
    raise OSError("handle is closed")
Exception in thread QueueManagerThread:
Traceback (most recent call last):
  File "/usr/lib/python3.7/threading.py", line 917, in _bootstrap_inner
    self.run()
  File "/usr/lib/python3.7/threading.py", line 865, in run
    self._target(*self._args, **self._kwargs)
  File "/usr/lib/python3.7/concurrent/futures/process.py", line 368, in _queue_management_worker
    thread_wakeup.clear()
  File "/usr/lib/python3.7/concurrent/futures/process.py", line 92, in clear
    while self._reader.poll():
  File "/usr/lib/python3.7/multiprocessing/connection.py", line 255, in poll
    self._check_closed()
  File "/usr/lib/python3.7/multiprocessing/connection.py", line 136, in _check_closed
    raise OSError("handle is closed")
OSError: handle is closed
OSError: handle is closed

[brianquinlan]

I really wanted ThreadPoolExecutor and ProcessPoolExecutor to have the same API when I designed them.

Do you have any bandwidth to debug this? If not, I could take a look.

[hniksic]

Please do take a look if you can. I am not acquainted with the implementation of ProcessPoolExecutor, so it would take quite some time for me to trace what's going on.

It would of course be ideal if both classes supported the new flag.

[brianquinlan]

I was playing with ProcessPoolExecutor and it seems like there are a bunch of problems that are triggered when pool.shutdown(wait=False) is used. I filed a bug for one issue: https://bugs.python.org/issue39205

Do you think that your PR could hold off until I have a chance to sort some of this out?

[hniksic]

Do you think that your PR could hold off until I have a chance to sort some of this out?

Sure, thanks for asking. We have a workaround, so it's no problem to wait for the proper solution. It's just that the workaround is so extremely ugly, involving monkey patch of a private method, that we'd definitely prefer the proper fix to land eventually.

[github-actions]

This PR is stale because it has been open for 30 days with no activity.

[hniksic]

I've now updated this PR to address the feedback from the issue discussion. The wait_at_exit parameter has moved from shutdown() to the constructor as a daemon=False keyword-only parameter, as agreed upon in the issue. The parameter is on the constructor only, keeping the shutdown() API consistent across executor types (per @brianquinlan's review comment). As before, shutdown(wait=True) still joins threads explicitly.

Since Python 3.9 made executor threads non-daemon, the implementation now also creates actual daemon threads, not just removes them from _threads_queues.

Tests are updated accordingly.

cc @brianquinlan