A partition can have three assigned replicas while only two are keeping up with new records. The third broker might still be running, answering requests, and storing older data. Counting assigned copies alone does not tell us which replicas can support the partition’s current replication guarantees.
Kafka tracks that distinction through the in-sync replica set, usually called the ISR. We’ll examine how membership changes, what those changes mean for writes and reads, and how to recognize a partition that is losing replication protection.
The examples use Apache Kafka 4.3 with KRaft, local broker storage, and non-transactional records. Assume the cluster enables Eligible Leader Replicas, or ELR. This is the default for newly created clusters starting with Kafka 4.1, but an upgraded cluster’s feature state can differ.