advanced

Kafka replication

Балансируйте replication factor, in-sync replicas, acknowledgements и leader election, trading throughput for durability and availability.

У каждой partition есть leader и followers, реплицирующие log. Replication factor и min.insync.replicas задают, сколько копий должны acknowledge write до commit. Leader election продвигает follower при падении leader.

Компромисс: выше replication и строже acks — лучше durability, но больше latency и load на brokers.

На интервью: acks=all с min.insync.replicas=2 на three-broker cluster; риск unclean leader election.

Типовые ошибки: replication factor 1 в production; acks=1 теряет data при leader crash; ISR shrink без alerting.

Чеклист:

RF и min.insync.replicas под durability goals.
Under-replicated и offline partitions в мониторинге.
Drills broker failure и leader election.
Producer ack configuration per topic class.