Skip to content

Poor collection performance due to expensive iscoroutinefunction check #720

Closed
@etripier

Description

@etripier

Hello pytest-asyncio developers! We've been using your plugin for a while. So much so that we now have tens of thousands of tests and hundreds of thousands of fixture arguments and ... it takes over three minutes to collect our tests.

Of that time, roughly two and a half minutes is taken up by pytest-asyncio. Specifically, this piece of code:

if fixturedef in processed_fixturedefs or not _is_coroutine_or_asyncgen(
.

That ends up calling https://github.com/python/cpython/blob/3aea6c4823e90172c9bc36cd20dc51b295d8a3c4/Lib/inspect.py#L384, which unwraps each fixture argument, makes sure it's a function, and then finally checks its co_flags to see if it's been marked as a coroutine function. The thing is, because you're feeding fixtures through, they tend to all be functions.

We ended up forking the plugin and inlining iscoroutinefunction and isasyncgenfunction - just the bit masking lines. Then we removed the code that was attaching the scoped_event_loop fixture because we don't use it and it was compounding the issue described above. That took us down to one minute.

I think it would help a lot to make the same changes here upstream. You could also swap the ordering of the comparison with

if not _is_asyncio_fixture_function(func) and asyncio_mode == Mode.STRICT:
, which would allow strict mode users to shortcut the expensive hot path. Finally, making scoped_event_loop optional would bring your plugin even closer to baseline performance.

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions