Decouple producers from consumers and smooth traffic spikes.

Data Flow

producer --> broker --> consumer(s)

Kafka (log-based):     partition 0 .. partition 1 .. partition 2
                       each event has offset; consumers track their own

RabbitMQ / SQS (queue-based):
                       producer -> exchange -> queue -> consumer
                       message removed after ACK

Two models

• Point-to-point (queue): message processed by exactly ONE consumer (SQS, RabbitMQ).
• Pub-sub (topic): message broadcast to ALL subscribers (Kafka topics, Google Pub/Sub).

Pitfalls

• Poison messages: a bad message crashes every consumer in rotation. Use DLQ (dead-letter queue).
• No idempotency: retries cause duplicates. Make consumers idempotent.
• Backpressure: the consumer’s slow — the queue grows unbounded. Set quotas.

Analogy

A mailbox at a post office — one slot holds the next letter until the postman picks it up.

Interview tip: Specify delivery semantics — at-most-once, at-least-once, exactly-once — and how you achieve them.