What happens when a Kafka broker goes down?

In a KRaft cluster, brokers send periodic heartbeats to the controller; a broker that misses them past broker.session.timeout.ms (default 9000 ms) is considered offline. If it was a follower, it loses its session and is removed from the ISR. For partitions where it was the leader, the controller elects a new leader from the remaining members of the ISR. Any Kafka node can answer metadata requests about live servers and partition leaders, so producers and consumers learn the new leaders from fresh metadata and send their requests there. The controller is a separate role from a broker in KRaft; if the controller itself fails, another controller is elected. When the failed broker restarts, it must catch up from the leaders and fully resync before it rejoins the ISR. If every replica of a partition fails, by default Kafka waits for an in-sync replica to return and favors consistency, so the partition may stay unavailable.

Expected key concepts: controller, heartbeat, broker.session.timeout.ms, new leader, ISR, elects, metadata, follower, removed, restarts, catch up