Details
-
Bug
-
Status: Resolved
-
Critical
-
Resolution: Fixed
-
3.6.0
-
None
Description
KAFKA-15021 introduced a new controller behavior that avoids increasing the leader epoch during the controlled shutdown scenario. This prevents some unnecessary thrashing of metadata and threads on the brokers and clients.
While a cluster is in a KIP-866 migration and has a KRaft controller with ZK brokers, we cannot employ this leader epoch bump avoidance. The ZK brokers must have the leader epoch bump in order for ReplicaManager to react to the LeaderAndIsrRequest.