Test suite flakes under pytest -n auto: cleanup_asset_dir wipes an asset directory shared by all xdist workers

[mickzijdel]

Description

src/anthias_server/api/tests/test_v1_endpoints.py fails intermittently under pytest -n auto, which is how CI runs the suite. It passes 100% of the time serially, so the tests themselves are correct — the failures come from xdist workers sharing one directory.

Measured on a local run, same commit, no code changes between runs:

$ for i in 1 2 3 4; do pytest src/anthias_server/api/tests/test_v1_endpoints.py -n auto -q | tail -1; done
28 passed in 4.20s
1 failed, 27 passed in 3.75s
2 failed, 26 passed in 3.41s
1 failed, 27 passed, 1 error in 3.14s

$ pytest src/anthias_server/api/tests/test_v1_endpoints.py -q | tail -1
28 passed in 0.93s

Two failure signatures, both from the same cause:

E  FileNotFoundError: [Errno 2] No such file or directory: '<assetdir>/cb7013dbb18b4826afc4dabc794f04b8.tmp'
E  AssertionError: assert b'AAAA' == b'AAAABBBB'

The first is a staging file deleted between the upload creating it and the test reading it. The second is a chunked upload whose first chunk was removed before the second landed, so the reassembled file is short.

Cause

The cleanup_asset_dir fixture empties the real, shared asset directory in teardown:

@pytest.fixture
def cleanup_asset_dir() -> Iterator[None]:
    try:
        yield
    finally:
        asset_directory_path = Path(anthias_settings['assetdir'])
        for file in asset_directory_path.iterdir():
            file.unlink()

Every test in the module requests it, and settings['assetdir'] is one path shared by every xdist worker in the process pool. So whenever two upload tests overlap, the one that finishes first unlinks the other's in-flight staging files. Nothing is wrong with the assertions; they are observing another worker's teardown.

This is timing-dependent, so it will surface on CI as an unrelated red build on an innocent PR — the most expensive kind of flake, because the natural response is to re-run rather than investigate.

Proposed fix

Give each test its own directory instead of cleaning a shared one. Three sibling modules in the same package already do exactly this:

monkeypatch.setitem(settings, 'assetdir', str(tmp_path))

(test_xibo_import.py, test_screencloud_import.py, test_pisignage_import.py)

Applied here, cleanup_asset_dir points assetdir at a per-test tmp_path and drops the unlink loop entirely — pytest removes the directory itself. That makes the isolation the tests already assume real, removes the teardown, and keeps them working under any -n value.

Worth a look afterwards at whether any other suite mutates shared on-disk or Redis state the same way, since the same pattern would flake for the same reason.

Happy to send a PR for this if it looks right.