RabbitMQ — Common Interview Questions and Answers

1. What is RabbitMQ and when do you use it?

RabbitMQ is a message broker implementing AMQP 0.9.1 (with extensions). It accepts messages from producers, routes them through exchanges to queues, and delivers to consumers.

Use it for:

• Work queues (background jobs, async processing).
• Pub/sub with multiple consumers.
• RPC-style request/response.
• Decoupling services with complex routing needs.

It’s not optimized for:

• Very high throughput streaming (Kafka wins).
• Long-term log retention with replay (Kafka wins).

2. Producer → Exchange → Queue → Consumer — explain.

Producers don’t send to queues directly; they publish to exchanges with a routing key. The exchange routes to bound queues based on its type:

• direct — exact routing key match.
• topic — pattern match (orders.*.placed, #.failed).
• fanout — broadcast to all bound queues.
• headers — match by header values.

A queue gets messages only if it’s bound to an exchange with a binding rule that matches the routing key.

3. What’s an exchange binding?

A binding is a routing rule attaching a queue to an exchange.

channel.queue_bind(exchange="orders", queue="email", routing_key="order.placed")

The exchange consults bindings on each publish: “for this routing key, which queues should I deliver to?” One queue can be bound to multiple exchanges; one exchange can have many queues bound.

4. Classic vs Quorum vs Stream queues?

For new deployments: quorum queues for tasks, streams for streaming. Avoid classic mirrored queues.

5. How do you make messages durable?

Three things, all required:

1. Durable queue — durable=True on queue_declare. Queue survives broker restart.
2. Persistent messages — properties=BasicProperties(delivery_mode=2). Messages written to disk.
3. Publisher confirms — channel.confirm_delivery() + check return value. Broker acknowledges receipt.

Even with all three, you have at-least-once delivery, not exactly-once. Consumer must be idempotent.

6. What’s a DLX and when do you use it?

Dead-Letter Exchange — set x-dead-letter-exchange on a queue; rejected/expired/over-limit messages route there.

channel.queue_declare(queue="orders", arguments={
    "x-dead-letter-exchange": "dlx",
    "x-dead-letter-routing-key": "failed-orders",
})

Use cases:

• Poison messages that fail every retry → DLX → DLQ → human investigation.
• Time-bounded retries (TTL on retry queue → DLX back to main queue).
• Failed Celery tasks (Celery + RabbitMQ + DLX pattern).

7. What’s the prefetch count and how do you tune it?

basic.qos(prefetch_count=N) — how many unacked messages the broker pushes to a consumer at once.

• Too low (1): every message requires a round-trip; throughput suffers.
• Too high (1000): one slow consumer hogs a thousand messages; the rest sit idle.
• Typical: 10-100 for short work, 1 for long-running tasks.

For Celery: worker_prefetch_multiplier defaults to 4 × worker_concurrency. Lower it for slow tasks to avoid starvation.

8. What’s the difference between basic.ack, basic.nack, and basic.reject?

• ack — message processed successfully; broker can delete it.
• nack — negative ack with multiple and requeue options. With requeue=False, message is dead-lettered (or dropped if no DLX). nack can batch (multiple=True).
• reject — older API, single message, requeue option.

Default Celery behavior: ack on success, nack with requeue=False on hard failure → DLX picks it up.

9. RabbitMQ vs Kafka — when each?

Most companies end up running both: RabbitMQ for Celery tasks, Kafka for cross-service event streaming.

10. How does RabbitMQ achieve HA?

Quorum queues replicate across N nodes via Raft consensus. A queue with N=5 tolerates 2 node failures. Writes go to the leader; replicated to a majority of followers before ack.

For broker discovery: a cluster of RabbitMQ nodes federates via Erlang networking. Clients connect via a load balancer that distributes across nodes (or use rabbitmq-uri with multiple hosts).

Classic mirrored queues (deprecated) replicated to designated followers; less reliable failover semantics than quorum.

11. What’s a publisher confirm?

By default basic.publish doesn’t return anything — fire-and-forget. The broker may not have actually persisted the message.

channel.confirm_delivery()
ok = channel.basic_publish(
    exchange="...", routing_key="...", body=...,
    mandatory=True,
)
if not ok:
    # broker didn't confirm; resend or alert
    ...

Confirms add latency (broker waits to confirm) but guarantee durability. For mission-critical work, always enable confirms.

12. What’s the mandatory flag?

mandatory=True on publish: if no queue is bound to the exchange with a matching routing key, the message is returned to the publisher (not silently dropped).

Useful for catching routing-key typos and dead consumers. The Channel must have a basic_return callback registered to receive the bounce.

13. Why are channels not threads-safe?

The AMQP wire protocol is multiplexed across one TCP connection via channels — each channel is a logical session. A single channel processes commands serially; concurrent thread access garbles the stream.

Pattern: one connection per app, one channel per thread/coroutine. Pools manage this.

14. What’s a lazy queue?

A queue that prefers disk over RAM. Default queues cache messages in RAM and only write to disk for durability; under deep backlog, the broker OOMs.

channel.queue_declare(queue="bulk", arguments={"x-queue-mode": "lazy"})

Use for queues that may accumulate large backlogs of large messages. Slightly slower per-message but won’t blow up the broker.

In modern RabbitMQ (3.12+), quorum queues have similar storage semantics and lazy queues are less common.

15. How do you handle a poison message?

A poison message fails every attempt — malformed payload, missing reference, etc.

Pattern:

1. Consumer fails, nack with requeue=False.
2. Configured DLX routes the message to a DLQ.
3. Alert on DLQ depth growth.
4. Human inspects: fix the bug, or delete the message if unfixable.
5. Replay the DLQ (shovel back to main queue) once fixed.

Without DLX, a poison message either bounces forever (if requeue=True) or vanishes silently. Always configure DLX for production queues.

16. What’s RabbitMQ’s “shovel”?

A built-in plugin that copies messages from one queue/exchange to another — possibly on a different broker. Used for:

• Cross-data-center replication of specific queues.
• Replaying a DLQ back to the main queue.
• Migrating queues between brokers.

rabbitmq-plugins enable rabbitmq_shovel

Configure via management UI or rabbitmq.conf.

17. RabbitMQ vs Celery — what’s the relationship?

Celery is a task queue library; RabbitMQ is one of its supported brokers (Redis, SQS, others also work). Celery uses RabbitMQ exchanges + queues internally:

• Each task queue is a RabbitMQ queue.
• Routing keys distinguish task names.
• DLX configuration on the Celery queue enables the DLQ pattern.

Many “Celery with RabbitMQ” production issues are actually RabbitMQ tuning issues — prefetch, durability, DLX.

18. Can you do RPC with RabbitMQ?

Yes — the classic AMQP RPC pattern:

1. Client declares an exclusive temporary reply queue.
2. Client sends request with reply_to (the temp queue) and a correlation ID.
3. Worker processes, publishes reply to the reply_to queue with the correlation ID.
4. Client consumes from the temp queue, matches by correlation ID.

result = channel.queue_declare(queue="", exclusive=True)
reply_queue = result.method.queue
corr_id = str(uuid.uuid4())

channel.basic_publish(
    exchange="", routing_key="rpc_queue",
    properties=pika.BasicProperties(reply_to=reply_queue, correlation_id=corr_id),
    body=request,
)

# Consume from reply_queue, match correlation_id

Rarely the right tool — HTTP/gRPC is usually clearer. Use RabbitMQ RPC when you want broker-mediated load distribution + queue-based backpressure.

19. How does message TTL work?

Two scopes:

• Per-queue: x-message-ttl on queue declare. Every message expires after N ms.
• Per-message: expiration on publish, as a string of ms.

Expired messages are dropped (or dead-lettered if DLX configured). Useful for:

• Cache invalidations that go stale.
• Time-sensitive notifications.
• Retry-with-delay patterns (TTL on retry queue → DLX back to original).

20. Production tips?

• Always durable=True and delivery_mode=2 for production.
• Use quorum queues, not classic mirrored.
• Configure DLX on every operational queue.
• Set basic.qos(prefetch_count=...) consciously per consumer.
• Publisher confirms for important publishes.
• Monitor: queue depth, message rate, unacked count, connection count.
• Connection per app, channel per thread.
• Don’t rely on memory queues for large backlogs — quorum or lazy.

Mental model