New event loop is created across different runs in asyncio

I think this is not intentional - the implication is that setting new event loop policies will possibly lose the current event loop:

https://github.com/agronholm/anyio/blob/14daa2bad967bf6dd0f96d04ecffe8e383985195/anyio/_backends/_asyncio.py#L71-L86

This causes the pytest plugin to have different event loops across fixtures and tests:

import asyncio

import pytest


@pytest.fixture
async def loop():
    return asyncio.get_event_loop()


@pytest.mark.anyio
async def test(loop):
    assert loop is asyncio.get_event_loop()  # fails here

In many cases, the fixture should share the same loop as the test, for example:

import asyncpg

import pytest


@pytest.fixture
async def conn():
    return await asyncpg.connect("postgresql:///")


@pytest.mark.anyio
async def test(conn):
    assert await conn.fetchval("SELECT 123") == 123

E   RuntimeError: Task <Task pending name='Task-5' coro=<run.<locals>.wrapper() running at .../anyio/_backends/_asyncio.py:67> cb=[run_until_complete.<locals>.<lambda>()]> got Future <Future pending cb=[Protocol._on_waiter_completed()]> attached to a different loop

One possible solution is to cache all policy instances under different setups, and switch to use the right one. I’ll create a PR with a more complete test.

[agronholm]

I have a plan for this: run async fixtures in a separate thread where the event loop will be kept running. The tricky part is figuring out when to start and stop the event loop. Each back-end needs one thread each. This might also make it feasible to allow class/module/package level async fixtures.

[agronholm]

I just pushed the fully refactored pytest plugin to the testrunner branch. Feel free to check it out!