Parametrizing `event_loop_policy` parametrizes all tests

[aceg1k]

Contrary to the statements in the documentation (here and here), the fixture is not only applied to all pytest-asyncio tests, but also applied to all other tests.

import asyncio
import pytest
import uvloop

@pytest.fixture(
    scope="session",
    params=(asyncio.get_event_loop_policy(), uvloop.EventLoopPolicy()),
    ids=("asyncio", "uvloop"),
)
def event_loop_policy(request):
    return request.param

@pytest.mark.asyncio
async def test_async():
    pass

def test_sync():
    pass

Output:

plugins: asyncio-0.23.5
asyncio: mode=Mode.STRICT
collected 4 items                                                                                                       

test_event_loop_policy.py::test_async[asyncio] PASSED                                                             [ 25%]
test_event_loop_policy.py::test_sync[asyncio] PASSED                                                              [ 50%]
test_event_loop_policy.py::test_async[uvloop] PASSED                                                              [ 75%]
test_event_loop_policy.py::test_sync[uvloop] PASSED                                                               [100%]

[seifertm]

Thanks for the report and the reproducer!

The event loop policy fixture is defined as an autouse fixture. I don't exactly recall why, but I think the reason is that it's now used by the deprecated event_loop fixture and had to be marked as autouse for backwards compatibility.

It's obviously a bug that this also parametrizes all sync tests and needs to be addressed.

However, I don't expect this issue to be fixed before v1.0 when the legacy event_loop fixture is gone.

[sean-anderson-seco]

Ran into this today. Unfortunately, this means that event_loop_policy is useless for me. So the easiest way is to go back to event_loop:

import asyncio
import contextlib
import warnings

import pytest
import uvloop

@pytest.fixture(
    scope="session",
    params=(asyncio.get_event_loop_policy(), uvloop.EventLoopPolicy()),
    ids=("asyncio", "uvloop"),
)
def policy(request):
    warnings.filterwarnings('ignore', "The event_loop fixture")
    return request.param

@pytest.fixture
def event_loop(policy):
    with contextlib.closing(policy.new_event_loop()) as loop:
        yield loop

@pytest.mark.asyncio
async def test_async(policy):
    pass

def test_sync():
    pass

Note that you must parametrize your async tests by policy in some way, otherwise you will get The requested fixture has no parameter defined for test. This is because event_loop is requested using request.getfixturevalue, so pytest doesn't know that it should create two runs of test_async, and when event_loop is requested it is too late. In my case I was already requesting policy in all my tests anyway, but this may be more annoying for you.

[tjkuson]

Having this work with async test items collected under the auto mode will be tricky as it'd mean finding a way to parameterize tests after they've been collected (encountered this issue before in #810 (comment)).

[seifertm]

Having this work with async test items collected under the auto mode will be tricky as it'd mean finding a way to parameterize tests after they've been collected (encountered this issue before in #810 (comment)).

Given this problem and with #1454 closed, I think the right path forward is a timely removal of the policy system in pytest-asyncio. Users can stay on v1.4.0 to upgrade from the policy system to the hook-based approach for the time being.