Kafka brokers

A broker is a Kafka server process. A cluster is a set of brokers, and the data (i.e. partitions) is spread across them. Each partition has one leader broker (handles reads and writes for that partition) and followers (replicas) on other brokers that mirror it. Clients connect to any broker ("bootstrap server"), get metadata about who leads what, then talk directly to leaders for actual IO. When a broker dies, one of its follower replicas is promoted to leader and service continues. Two things to internalise: brokers don't hold application state — they only hold logs, so you can add/remove brokers without losing data as long as replication is healthy; the bottleneck is usually disk and network, not CPU — throughput is about sequential disk writes and network bandwidth, which is why Kafka is fast on commodity hardware. Historically ZooKeeper managed cluster metadata; modern Kafka (KRaft mode) runs that itself via Raft.