mirror of
https://github.com/maziggy/bambuddy.git
synced 2026-09-30 03:01:21 +02:00
maziggy
e0a2398783
Backend Tests (shard 1/4) timed out after 10 minutes on e8b901f54
("Updated BACKERS"), a docs-only commit. The step was not hung: it
reached 98%, every test passing, and was killed about 12 seconds short
of finishing.
pytest-split balances by duration only when it has a durations file,
and there was none. Without one it splits by test COUNT -- 2926 / 2926
/ 2926 / 2925, exactly 25% each, which is what the comment here claimed
was good enough. Count is not time. Measured over the full suite
(11703 tests, 1045.6s):
shard 1 2926 tests 658.2s
shard 2 2926 tests 173.1s
shard 3 2926 tests 113.3s
shard 4 2925 tests 101.1s
Shard 1 was carrying 63% of the suite's runtime -- a 6.5x spread -- and
had been walking toward the cap for weeks: 475s, 404s, 445s, then 587s
on 08-30, thirteen seconds under it, and 613s here. CI printed the
reason in its own log on every run: "[pytest-split] No test durations
found."
Commit tests/.test_durations and pass --durations-path explicitly:
shard 1 1273 tests 261.6s
shard 2 1070 tests 261.5s
shard 3 1522 tests 262.0s
shard 4 7838 tests 260.6s
A 1.01x spread. The lopsided test counts are the point: shard 4 takes
thousands of fast unit tests, shard 1 keeps the slow integration ones.
The four groups still partition the suite exactly -- union is 11703
with nothing dropped or duplicated, and all four pass.
--durations-path has to be explicit because pytest-split defaults it to
$CWD/.test_durations, and the two matrices run from different
directories: the native job from backend/, the Docker job from the
image root. Left implicit, the Docker one finds nothing and silently
falls back to the count split. Node IDs match across both because
backend/tests/pytest.ini pins rootdir to backend/tests either way, so a
single file serves them both; verified the file clears .dockerignore
and lands at /app/backend/tests/.test_durations at full size.
timeout-minutes 10 -> 15 for headroom, since the file goes stale as
tests are added. Staleness degrades slowly rather than breaking --
unknown tests are treated as average.