Provide environment information

• Self-hosted Trigger.dev v4.8.3 (Docker Compose), linux/arm64
• External PostgreSQL reached through a TCP proxy, Redis 7
• The same code is in v4.8.5 and on main as of 2026-11-12

Describe the bug

A short PostgreSQL outage permanently stops cron schedules. Ours lasted about 35 seconds and stopped 8 of 123 active schedules. They never fired again until we called the admin recover endpoint (/admin/api/v1/environments/{id}/schedules/recover). Nothing in the dashboard shows that a schedule has stopped.

Why it happens:

1. Each tick is one schedule-engine worker job, schedule.triggerScheduledTask. The next tick is only registered at the end of a successful handler (ScheduleEngine.triggerScheduledTask → registerNextTaskScheduleInstance).
2. scheduleWorkerCatalog gives that job retry: { maxAttempts: 5 } (internal-packages/schedule-engine/src/engine/workerCatalog.ts). With the redis-worker defaults (factor 2, 1 s base, randomized) all five attempts finish in about 15–30 seconds.
3. If PostgreSQL is unreachable for that whole window, the job goes to the dead-letter queue. No job exists for the instance any more, so nothing registers the next tick. The chain is broken for good.
4. The handler is called with finalAttempt: false, // TODO: implement retry logic (internal-packages/schedule-engine/src/engine/index.ts), so there is no last-attempt fallback.

Reproduction repo

No repo needed; this reproduces on a stock self-hosted install.

To reproduce

1. Self-host v4.8.3 with a few * * * * * schedules in a production environment.
2. Stop PostgreSQL for about 75 seconds, starting just before a minute boundary. Then start it again.
3. Check Redis. Most schedules now have no entry in schedule:schedule:{queue:schedule-engine-worker:}items and one in …dlq. They never fire again.

In an isolated test with four schedules, three stopped permanently after a 76 second outage.

Suggested fix

• Reconcile periodically. Run the existing recoverSchedulesInEnvironment logic on a timer for all active, non-development instances, using enqueueOnce so a live job is never replaced. This makes the chain self-healing whatever the cause. We run a 60 second loop as a local patch; with it, a 12 minute outage recovered six seconds after PostgreSQL returned, with no duplicate runs.
• A longer retry budget for this job, for example maxAttempts: 15 with maxTimeoutInMs: 60_000 (about 9–10 minutes). A short outage then costs a late tick instead of a lost schedule. calculateNextSchedulableOccurrence already skips expired slots, so there is no backlog replay.
• An idempotency key on the scheduled trigger (instance id + schedule time). Today a retry after "trigger succeeded, next-tick registration failed" can trigger the same slot twice, and a longer retry budget makes that more likely.
• Make it visible. A metric or dashboard state for "active schedule with no pending job" would surface this.

Additional information

The webapp log at the time showed Failed to trigger scheduled task and then Failed to register schedule instance, both with Prisma Can't reach database server, for each affected instance. The dead-letter entries carry the same error with attempt: 4.