Should event_loop fixture call asyncio.set_event_loop()?

[adamrothman]

I use pytest and pytest-asyncio to test a few Tornado applications. I had some trouble (see tornadoweb/tornado#2324) with Tornado getting a different event loop from the one created by pytest-asyncio's event_loop fixture.

This can be fixed by overriding Tornado's default testing behavior to use the "current" loop, rather than creating a new one. The issue is that pytest-asyncio does not call asyncio.set_event_loop with the loop it creates in the event_loop fixture – without which, there is no "current" loop.

I'd like to propose change the event_loop fixture to look like this:

@fixture
def event_loop(request):
    loop = asyncio.new_event_loop()
    asyncio.set_event_loop(loop)
    yield loop
    loop.close()
    asyncio.set_event_loop(None)

How does that sound? I'm happy to submit a PR, but I wanted to get feedback on the idea first.

[Tinche]

Hm, I think it used to be considered bad form to call asyncio.get_event_loop before, but now it's considered OK, so we might enable this. @asvetlov thoughts?

In any case, we use a fixture hook to do things like this in order to keep the event_loop fixture very simple, so it's simple to override.

[asvetlov]

It is safe starting from Python 3.5.3

[asvetlov]

pytest-asyncio operates with async tests only.
The async test always have an event loop available by both asyncio.get_event_loop() and asyncio.get_running_loop() calls.
There is no need to explicitly call asyncio.set_event_loop(), it is done by loop.run_untile_complete() already.