Release 0.14 breaks tests that rely on a specific event_loop_policy

[wouterdb]

Problem

We have many testcases that rely on asyncio.set_event_loop_policy(AnyThreadEventLoopPolicy())
We set the event loop policy in the conftest.py, such that it is set before any tests start.

Since 0.14 these testcases produce errors

RuntimeError: There is no current event loop in thread 'ThreadPoolExecutor-109_6'.

Root cause

The change here: https://github.com/pytest-dev/pytest-asyncio/pull/164/files#diff-4d85e8313d8e67cb777137411a11ac86R80

Resets the event loop policy after the first test, breaking all the next tests.

[Tinche]

@wouterdb will dig into this over the weekend!

[wouterdb]

Thx!

…

On Sat, Jun 27, 2020, 03:45 Tin Tvrtković ***@***.***> wrote: @wouterdb <https://github.com/wouterdb> will dig into this over the weekend! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#168 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANUSXUE36Y2RYD7VUFIFDDRYVFKJANCNFSM4OIKB57Q> .

[alblasco]

Thanks also for your kind feedback. I did V0.14 changes you are referring, and I was not fully aware of your user case.

With this in mind and after some debugging, I see no problem to roll back some of V0.14 changes:
a) Restore old_loop = policy.get_event_loop() and its finalizer fixturedef.addfinalizer(lambda: policy.set_event_loop(old_loop))
https://github.com/pytest-dev/pytest-asyncio/pull/164/files#diff-4d85e8313d8e67cb777137411a11ac86L58-L65
b) Remove finalizer: asyncio.set_event_loop_policy(None) https://github.com/pytest-dev/pytest-asyncio/pull/164/files#diff-4d85e8313d8e67cb777137411a11ac86R75-R80

@Tinche With these changes all our unitary tests keep passing.
BTW. It seems that skipping V0.13, has not been enough to avoid issues.

[Tinche]

Yeah, it would be great if we could treat the loop policy in the plugin as read-only (i.e. use it for generating event loops but let users override it). Another use case I imagine is very common is testing with UVLoop.

[Tinche]

@alblasco I've done some work on the 'policies' branch, but I still have trouble making all the tests pass. The first test in a package still uses the wrong event loop for some reason.

[Tinche]

@wouterdb I'm doing some work to support this use case over at #174.

Due to a lot of users using the plugin in a multitude of ways, it's not super trivial to fix. The fix I'm proposing for now is to treat fixtures named event_loop_policy a special way (like we treat fixtures called event_loop a special way). Then we can ensure those fixtures are executed at the right times. Would this work for you? I guess you would have to set the policy using a fixture and that fixture needs to be named event_loop_policy.

[wouterdb]

This would work for us. (It would actually be a lot cleaner than what we do now.)

Thank you for you fast response.

[bryevdv]

I guess you would have to set the policy using a fixture and that fixture needs to be named event_loop_policy.

I'm afraid I don't follow all the discussion, but this sentence seems troublesome. In our case the library itself is setting a policy (because we utilize tornado which cannot use the default proactor on windows). I maintain that we should not have to do or set anything at all in our tests in order to have the library-configured policy be respected as-is (otherwise, we fail to be able to test that the library itself is configuring the policy, as it must do).

[wouterdb]

we should not have to do or set anything at all in our tests in order to have the library-configured policy be respected as-is

To us, it would be good if pytest doesn't touch the policy (as it did before), but if it does reset it, I would prefer to be able to control what it is reset to.

The problem for us is that pytest now resets the event_loop_policy after every test.
Many of our unit tests require an AnyThreadEventLoopPolicy. Many of our fixtures are asynchronous as well.
So by the time the actual test starts, the event loop is already in use. Changing the event_loop_policy when the event loop is already in use is very messy and causes complicated errors:

(otherwise, we fail to be able to test that the library itself is configuring the policy, as it must do).

The reverse side of this is that we have many unit test that don't configure the policy themselves (the code that configures the policy is out-of-scope for that test) so we need to be able to define a fixture that sets the policy before any of these test are executed.

Current work around

def patch_policy():
    # work around for https://github.com/pytest-dev/pytest-asyncio/issues/168
    oldloop = asyncio.get_event_loop_policy()
    if not isinstance(oldloop, AnyThreadEventLoopPolicy):
        newloop = AnyThreadEventLoopPolicy()
        # transfer existing eventloop to the new policy
        newloop.set_event_loop(oldloop.get_event_loop())
        asyncio.set_event_loop_policy(newloop)