Repository navigation
Conversation
A job completion is only propagated through the in-memory jobEmitterQueue. When that emission is lost (crash or shutdown before the worker runs, or a report lost while the database is recovering), the run's dependent jobs stay waiting/blocked/pending forever and a restart does not heal them. Add ReEmitUnresolvedRuns, which re-derives the missing emission from run state: it re-emits runs whose latest attempt has a finished job but still has unresolved jobs. Register it as a cron task that runs at startup and every 5 minutes, so stalled runs recover without manual database surgery. Fixes: go-gitea#40279 Signed-off-by: Xiong Han <3324733895@qq.com>
GiteaBot
added
the
lgtm/need 2
github-actions
Bot
added
topic/gitea-actions
Can you explain how you "manually tested" since the code won't run at all ? |
RegisterTask requires an admin.dashboard.<name> translation for every cron task; without it the server fails to start (log.Fatal in RegisterTaskFatal during cron.Init), which broke all db/e2e tests on this PR. Add the missing entry and a regression test that registers the task. Signed-off-by: Xiong Han <3324733895@qq.com>
The new reemit_unresolved_action_runs cron task raises the total registered cron tasks from 30 to 31; update the admin cron list test assertions (X-Total-Count header and response length) to match. Signed-off-by: Xiong Han <3324733895@qq.com>
The List subtest asserts the total cron count (31) and the body length, but the default page size is 30, so the body was truncated while X-Total-Count reported 31. Request limit=50 (MaxResponseItems) so the body holds all tasks, and assert the new reemit_unresolved_action_runs task is listed. Signed-off-by: Xiong Han <3324733895@qq.com>
|
You're right — the code couldn't start at all, and it was my change that broke it. The new reemit_unresolved_action_runs cron task was registered without the admin.dashboard.reemit_unresolved_action_runs translation entry that RegisterTask requires, so RegisterTaskFatal called log.Fatal during cron.Init() and the server died on startup. That is why all db-tests and e2e failed at startup while backend and test-unit passed. To be precise about the "manually tested" claim: what I had actually run locally was the unit tests for ReEmitUnresolvedRuns and the services/cron package, which passed — but those never exercise server startup, so they never hit the missing translation. I had not run the integration tests locally before opening the PR, and that gap is what let this through. Fixed in 2e4eec3: added the missing locale entry, plus a regression test (TestRegisterReEmitUnresolvedActionRuns) that registers the task and fails without the entry. I verified locally that the server now boots (integration TestMain path) and the new test passes. The cron list test also needed updating: the new task brings the registered total to 31, and the test now requests /api/v1/admin/cron?limit=50 so the response body is not truncated by the default page size of 30 — the X-Total-Count header and the body were measuring different things before. All checks are green on the latest head. |
|
diff is too large imho and contains massive comments, trim it down please, ensure AGENTS.md is being followed. |
Shorten the ReEmitUnresolvedRuns doc comment and drop the paragraph comments in the sweep tests, following AGENTS.md. Use RemoveByTags to clean up the cron registration test instead of iterating jobs. Assisted-by: Muse
|
Trimmed in fa829e7, following AGENTS.md:
No behavior or test changes: the sweep logic, the 6 re-emit cases, the unblock test, and the registration test are all untouched. Verified locally: Assisted-by: Muse |
Job completions propagate only through the in-memory jobEmitterQueue, so when an emission is lost (a crash before the worker runs, or a report lost while the database is recovering), dependent jobs stay waiting, blocked or pending forever, even across restarts.
This adds a startup and periodic sweep that re-runs the per-run check for runs whose latest attempt has a finished job and a still-unresolved dependent. Runs awaiting approval and fully settled runs are skipped.
Fixes: #40279