Skip to main content

Job queue

Source of truth: Postgres​

DATABASE_URL=postgres://...

Workers claim jobs with FOR UPDATE SKIP LOCKED. There is no file-backed job queue — it does not survive multi-replica deploys.

Optional broker (wake-up only)​

For KEDA / faster fan-out:

STEW_QUEUE_BROKER=rabbitmq   # or nats | pulsar
RABBITMQ_URL=amqp://...

Jobs are still written to Postgres first; broker publish is best-effort.

Disaster recovery​

If the broker loses messages:

  1. Workers still process pending jobs via Postgres poll
  2. Platform operators can republish pending jobs to rehydrate broker depth:
GET  /v1/platform/queue
POST /v1/platform/queue/republish

UI: Settings → Platform ops → Job queue recovery

Compose helper: deploy/compose/docker-compose.queue.yml.