Queues

Queues decouple producers from slow consumers—smoothing spikes, retrying work, and isolating failure domains. Examples: SQS, RabbitMQ, BullMQ, Kafka (log-oriented).

| Concern | Practice | |---------|----------| | Concurrency | Limit parallel consumers per queue | | Retries | Exponential backoff + max attempts | | DLQ | Park poison messages for inspection | | Idempotency | Handlers safe on duplicate delivery |

					API → enqueue job → Worker pool → side effects (email, PDF, webhook)
				

Monitor **queue depth** and **age of oldest message**—lag is a scalability signal. Scale consumers when lag grows, not only when CPU is high.

On interviews: at-least-once vs exactly-once; ordering guarantees; backpressure when queue grows unbounded; idempotent payment handlers.

Common pitfalls: unbounded queue hiding overload; no DLQ; consumers faster than downstream DB can handle.

The trade-off is resilience and elasticity versus eventual consistency and operational monitoring.

Checklist:

• Bounded retries with DLQ.
• Idempotent consumers with dedupe keys.
• Alert on lag and oldest-message age.
• Match consumer concurrency to downstream capacity.