During cleanup, should pytest-asyncio cancel outstanding tasks, shut down async generators, and shut down the thread pool, like `asyncio.run()` does?

[SyntaxColoring]

Here's how pytest-asyncio implements the event_loop fixture:

pytest-asyncio/pytest_asyncio/plugin.py

Lines 224 to 229 in 6ec7647

	@pytest.fixture
	def event_loop(request):
	"""Create an instance of the default event loop for each test case."""
	loop = asyncio.get_event_loop_policy().new_event_loop()
	yield loop
	loop.close()

Note that loop.close() is the sole cleanup code.

Compare this with the implementation of asyncio.run():

    loop = events.new_event_loop()
    try:
        events.set_event_loop(loop)
        if debug is not None:
            loop.set_debug(debug)
        return loop.run_until_complete(main)
    finally:
        try:
            _cancel_all_tasks(loop)
            loop.run_until_complete(loop.shutdown_asyncgens())
            loop.run_until_complete(loop.shutdown_default_executor())
        finally:
            events.set_event_loop(None)
            loop.close()

Before calling loop.close(), asyncio.run() does this:

1. It attempts to cancel outstanding tasks with _cancel_all_tasks().
2. It calls loop.shutdown_asyncgens()
3. It calls loop.shutdown_default_executor().

(I think (2) and (3) correspond to what the documentation calls "finalizing asynchronous generators and closing the threadpool.")

I only have a tenuous grasp on these asyncio internals, but it seems like the event_loop fixture should do this, too? And so should any other place in pytest-asyncio where the event loop is shut down, if there are more.

[andyljones]

Echoing this - I just spent a while figuring out where the heck these 'async generator not closed' warnings were coming from in my test suite, when I couldn't reproduce the behaviour in prod.

[podusowski]

+1

Got a bunch of "Task was destroyed but it is pending!" because tasks weren't allowed to handle cancellation and finish.

[Tinche]

Yeah, this seems like a good change to make. I'll take a PR or implement this myself in the next couple of weeks.

[asvetlov]

Can be solved as a side-effect of #235