Skip to main content

Runbook

What healthy looks like​

  • GET /actuator/health is UP.
  • Kong route for this service returns a non-5xx for a known probe path.
  • If the service produces events, notification.event_outbox pending count is near zero and lag is under 30 seconds.

Failure modes​

  • Inbox row exists but no push: token invalid, preference suppressed push, or dead-letter rate. Category toggles can keep the inbox row and skip push.
  • Duplicate notifications: consumer retried without event-id dedupe. Inspect inbox state before replaying a stream.

Diagnose​

  1. Take the traceId from the client error body or X-Request-Id.
  2. In Loki, filter service and traceId. Confirm outcome and httpStatus.
  3. If the symptom is delayed side effects, check outbox pending for schema notification and the consumer group named in Events.
  4. Check Hikari pool exhaustion and Redis latency on the service dashboard before restarting.

Repair​

Restart the single service container or Gradle process. Do not truncate outbox, inbox, or stream data to "clear" an error. Replay dead letters from the relay admin API after the consumer bug is fixed.