Skip to main content
The durable scheduler is an independently deployable service. It calculates interval, calendar, and cron occurrences in an IANA timezone and persists an occurrence before a worker can claim it.

​
Duplicate prevention

PostgreSQL advisory locking elects one scheduler transaction at a time. Row locks and the (schedule_id, scheduled_for) constraint prevent a restart or overlapping tick from emitting the same occurrence twice. Schedules define daylight-saving, misfire, jitter, concurrency, and no-runner behavior. A spring-forward wall time that does not exist is not silently shifted. Bounded backfill is explicit, and an existing occurrence is not re-emitted.

​
Queue delivery

Events use at-least-once delivery. A claim carries a short hashed visibility token. An expired claim returns to the queue and adds immutable attempt history. Retries use exponential backoff with deterministic jitter; the event enters dead-letter state after its attempt bound or three identical failures.

​
Replay

Administrative replay is audited and creates a new queue row, correlation ID, and idempotency key while preserving the original logical event identity. Replay does not prove an external side effect is safe; node recovery policy still applies. Payloads are referenced through encrypted storage rather than stored inline. Queue rows carry workspace, workflow revision, routing requirements, expiry, attempts, and correlation identity.