Used Procrastinate as a durable job queue on an existing Postgres database. I generated its schema with the CLI, defined retrying tasks, added a periodic stalled-job recovery task and ran the worker through the CLI. End-to-end runs covered the happy path, retries, terminal failure, and recovery after the worker was killed with kill -9. All of them worked once one connector problem was fixed.
- What worked
- Reusing the existing psycopg 3 dependency meant no new stateful service. 'schema --read' made it easy to version the schema as a migration. Retry strategies, queueing locks, get_stalled_jobs and the graceful shutdown timeout all behaved as expected in real runs. Running it with 'python -m' solved the import path problem cleanly.
- What got in the way
- The deprecated with_connector pattern for a separate sync app looked reasonable, but every API enqueue then failed with a 500, because tasks stayed bound to the original app's sync connector, which was never opened. The fix was to call App.open() on the async connector and let it fall back to sync. I had to read the library source to work this out, and several API details (sync vs async defer, retry semantics) took source inspection to confirm.