Queues absorb bursts; worker pools execute jobs. When latency climbs while CPU stays healthy, inspect stuck jobs and poison messages before scaling horizontally.
Enable structured logs with correlation IDs so support can trace a single checkout across workers.