MSSQL's default READ COMMITTED makes reads wait on writers, so the
runner pickup deadlocks with concurrent claims, flaking
`TestCreateTaskForRunnerConcurrentClaim`.
- Enable `READ_COMMITTED_SNAPSHOT` on MSSQL so it reads like PostgreSQL
and MySQL
- Read the pickup cursor before claiming, a lost claim could skip
waiting jobs
- Add tests that fail without consistent READ COMMITTED
Performance: Writes on MSSQL now also store the previous row version in
tempdb, the same versioning cost PostgreSQL and MySQL always pay, and
Azure SQL enables it by default. Reads no longer block on writers, and a
32-runner pickup stress test ran 2.5x faster with it.
---------
Signed-off-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: Giteabot <teabot@gitea.io>
## Summary
Fixes two related bugs that cause jobs in large workflows (50+ parallel
jobs) to never get a runner assigned even though runners are free.
### Bug 1 — Concurrent runner race
When N runners all poll `FetchTask` with a stale `tasksVersion`
simultaneously, they all query the same waiting job list sorted by
`(updated, id)` and all pick **job #1**. Only one wins the `UPDATE WHERE
task_id=0` optimistic lock; the rest return empty-handed but still
receive `latestVersion` in the response. They then consider themselves
"up to date" and skip `PickTask` on every subsequent poll, leaving jobs
#2–50 permanently unassigned.
**Fix:** `CreateTaskForRunner` now iterates through all matching waiting
jobs. When the optimistic lock fails on job #1, it immediately tries job
#2, then #3, etc., each in its own independent transaction so a failed
attempt rolls back cleanly before the next candidate is tried.
`PickTask` no longer wraps this call in an outer `db.WithTx` (which
caused `halfCommitter` entanglement that prevented per-attempt
rollbacks).
### Bug 2 — Idle runner doesn't re-check after finishing a task
`tasks_version` only bumps when a job transitions **to** waiting (new
workflow triggered, blocked→unblocked). After a runner finishes its
current task it polls `FetchTask` with `tasksVersion == latestVersion`,
so the server skips `PickTask` entirely — the remaining 45 waiting jobs
are invisible to the now-idle runner.
**Fix:** Also call `IncreaseTaskVersion` in `UpdateRunJob` when a
(non-reusable-caller) job transitions to a **done** state. Idle runners
then see a version mismatch on their next poll and attempt `PickTask`,
picking up the remaining jobs.
