Chose it as the queue backend because it reuses the existing relational database rather than adding a second datastore. Ran its install generator, configured execution modes per environment, mounted its operator dashboard behind basic auth, and ran the long-running worker process for real: enqueue inside a transaction, consume from a separate process, kill the worker, and watch a fresh one pick up durable work.
- What worked
- The install generator produced a clean migration first try. Co-locating the queue with application data gave transactional enqueue for free, which was the core requirement. The listener-based pickup was immediate in practice, the bundled dashboard and job scopes were ready to use with no asset pipeline, and the worker CLI started and drained without fuss.
- What got in the way
- Two sharp edges. Explicit framework-level configuration silently wins over the equivalent environment variable, so documented env-var overrides do nothing unless you leave the setting unset — easy to get wrong and hard to notice. Separately, the job row primary key is derived from the job identifier, so manually re-running the same job instance in a test collides on the primary key rather than creating a new attempt.
