Production BullMQ queues stay reliable when concurrency, retries, and job idempotency are chosen together rather than tuned alone. Prefer a bounded worker concurrency that matches Redis and disk throughput for PDF work, keep attempts capped with a backoff that does not stampede the renderer, and store a stable job key so a redeliver cannot duplicate a finished render. Separate the ingest queue from the heavy render queue so a burst of uploads cannot starve in-flight jobs.
Ops prove-out after tightening queue settings: enqueue a synthetic burst at target concurrency and confirm active job count plateaus at the limit, then force a worker kill mid-job and confirm the same key resumes once without double output. Watch Redis memory, stalled-job rate, and render p99 for one deploy window — if stalled climbs while concurrency stays flat, the lock duration is shorter than the heaviest PDF path. Keep a canary queue with the prior settings beside the new path so you can tell a bad backoff from a real renderer regression.