Railway worker

The production worker service uses /railway.worker.json: build runs pnpm worker:check (loads the handler registry without connecting), and the service runs pnpm worker continuously with restartPolicyType: ON_FAILURE. There is no cron schedule and no scheduler service.

What the worker does

scripts/worker.mts wires lib/jobs/worker.ts to getRepositories():

1. Sweeps (every minute, in-process): pending in-app notifications whose deliver_at has passed become notify jobs, and scheduled posts that are due and were not yet handed to SocialBu become publish-post jobs. Sweep jobs use deterministic ids, so several replicas never double-enqueue. Hourly, lease rows older than seven days are purged.
2. Claims up to WORKER_CONCURRENCY (default 2) queued jobs whose run_at has passed, plus running jobs whose lease expired. A claim creates the row job_leases/<jobId>.<attempt>; a 409 means another worker won.
3. Runs the handler from lib/jobs/handlers.ts, renewing the lease every third of its length (default lease 2 minutes). Success marks the job succeeded; failures retry with exponential backoff until max_attempts, then become dead. PermanentJobError stops retries; RetryJobError picks the next attempt time.
4. Sleeps adaptively (1 s up to 20 s) when idle. POST /wake on WORKER_PORT (or PORT) cuts the sleep short; web code can call wakeWorker() (lib/jobs/wake.ts) after enqueueing, using WORKER_WAKE_URL on Railway private networking. GET /health reports liveness.

Handlers

Operating it

• pnpm worker:once runs one forced sweep and one claim batch, then exits.
• Required variables: APPWRITE_ENDPOINT, APPWRITE_PROJECT_ID, APPWRITE_API_KEY. LUMENCLIP_DATA_BACKEND must be unset or appwrite.
• Verify a deploy by checking the worker started log line and, after enqueuing work, job succeeded lines. Never print variable values.