Queues

What is the difference between at-most-once, at-least-once, and exactly-once message delivery? Which does Kafka use by default?

Design a system that processes 10 million events per day from IoT devices with minimal latency.

A Kafka consumer group has 3 consumers and 3 partitions. One consumer dies. What happens?

What is a dead letter queue? Describe a scenario where you would need one.

How would you implement backpressure in a system where producers are faster than consumers?

Acknowledging a message before processing is complete

Symptom: Messages are lost when the consumer crashes after acking but before finishing processing.

Not handling poison messages — one bad message blocks the entire queue

Symptom: Consumer crashes on one specific message, restarts, fetches the same message again, crashes again, in an infinite loop.

Consumer group with more consumers than partitions

Symptom: Some consumer instances in the group are idle and never receive any messages.

Publishing to a queue without a TTL on the message

Symptom: Queue depth grows unboundedly; old, irrelevant messages from hours ago are still being processed.

Synchronous publish blocking the hot path

Symptom: API endpoint latency spikes whenever the queue broker is slow or unavailable.

Message Queue Models: Point-to-Point vs Pub-Sub vs Log-Based

The queue model defines how messages are routed from producers to consumers and what guarantees are made around ordering, replayability, and fan-out.