mirror of
https://github.com/Joulenap/joulenap.git
synced 2026-08-11 13:21:43 +02:00
Catubba
5b39f9ae1a
The suite failed roughly one full run in five, differently each time. Three independent causes, not one: 1. The lifespan started `missed-backup-check` and `interrupted-run-alert` as daemon threads that nothing joined, and the missed-run check opens a database session. Every test that builds an app leaked one, so a thread routinely outlived its test and kept using the database while the next test tore the engine down and built its own. They are now joined on shutdown, which is also the right behaviour in production: a restart no longer abandons an alert halfway out. The join is bounded well under Docker's stop grace, so a black-holing notification channel still cannot hold the process open. The job service's queue worker stays unjoined on purpose - it may be mid-backup, and blocking shutdown on a running vzdump is worse. 2. `_ensure_ready()` initialised the database lazily when the session factory was missing, which built a schema at `paths.db_path()` - the real database - from whichever caller got there first. `create_all` reflects once and then issues CREATEs, so two of those at the same time collide with "table X already exists", the error the suite reported. It now raises instead, so a caller that outlived its setup fails loudly rather than racing. One test was relying on that lazy path and gained the fixture it always needed; it was the only one, i.e. the only test that had been writing to the real database. 3. The 409 came from a test helper, not from the database race. `_run_route` waited for the run row to leave RUNNING, but the row is finalised inside the cycle while the power-off, the notification, the single-run lock and the queue's current entry all still follow. A second run of the same route inside that window is correctly rejected as already queued. The helper now waits for the queue itself to go idle. Both races were reproduced before being fixed - the schema collision by racing threads through a probe with no initialisation, the 409 by widening the window with a temporary sleep, which made it fail every run - and both stop reproducing with the fixes in place. Two regression guards, each confirmed to fail on the pre-fix code: an autouse fixture that diffs the live threads around every test and fails the one that leaves any behind, and a test pinning that using the database before initialising it raises. 636 passed, 2 skipped; 10 consecutive full runs green, with the real database untouched.