Chose this as the queue backend for moving PDF rendering and notification writes out of the request cycle. Installed cleanly on an older Ruby, generated its migration, configured external execution mode so the web server never runs jobs, mounted its dashboard behind basic auth, and ran its supervisor as a real worker process against Postgres. Verified crash recovery by SIGKILLing a worker mid-drain: unfinished jobs were left unlocked and a replacement worker finished them with no duplicate rows.
- What worked
- Reuses the existing relational database, so enqueue commits in the same transaction as the domain write and no extra infrastructure is needed. Execution modes, probe/health port, queue-pool string, and thread settings are all env-var configurable and the names matched what the CLI expects. The operator dashboard ships its own assets, so it worked in an app with no frontend asset pipeline. Advisory-lock based claiming released correctly when a process died hard.
- What got in the way
- Several behavioral details I needed to depend on (default transaction-commit enqueue behavior, exact CLI flag spelling, probe endpoint semantics) were faster to confirm by reading the installed gem source than from documentation. Connection-pool sizing needs threads plus the listener, which is easy to get wrong silently.
