hachikuji commented on code in PR #14053: URL: https://github.com/apache/kafka/pull/14053#discussion_r1329117983
########## core/src/main/scala/kafka/cluster/Replica.scala: ########## @@ -98,14 +103,20 @@ class Replica(val brokerId: Int, val topicPartition: TopicPartition) extends Log * fetch request is always smaller than the leader's LEO, which can happen if small produce requests are received at * high frequency. */ - def updateFetchState( + def updateFetchStateOrThrow( followerFetchOffsetMetadata: LogOffsetMetadata, followerStartOffset: Long, followerFetchTimeMs: Long, leaderEndOffset: Long, brokerEpoch: Long ): Unit = { replicaState.updateAndGet { currentReplicaState => + // Fence the update if it provides a stale broker epoch. + if (verifyBrokerEpoch && brokerEpoch != -1 && currentReplicaState.brokerEpoch.exists(_ > brokerEpoch)) { Review Comment: Just putting this up for discussion. The source of truth for broker registrations is the metadata log. An alternative implementation might be to check the metadata cache. Then this would be an equality check and we would always ensure that eventually we get the right replica. The downside possibly is that it might delay convergence after broker restart I guess, but not sure if it is a real issue. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: jira-unsubscr...@kafka.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org