[KAFKA-13720] Few topic partitions remain under replicated after broker lose connectivity to zookeeper

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.7.1
• Fix Version/s: 3.1.0
• Component/s: controller
• Assignee: Unassigned
• Reporter: Dhirendra Singh
• Created: 09/Mar/22 12:07
• Updated: 20/Jun/22 03:07
• Resolved: 17/Jun/22 03:04

Description

Few topic partitions remain under replicated after broker lose connectivity to zookeeper.
It only happens when brokers lose connectivity to zookeeper and it results in change in active controller. Issue does not occur always but randomly.
Issue never occurs when there is no change in active controller when brokers lose connectivity to zookeeper.
Following error message i found in the log file.

[2022-02-28 04:01:20,217] WARN [Partition __consumer_offsets-4 broker=1] Controller failed to update ISR to PendingExpandIsr(isr=Set(1), newInSyncReplicaId=2) due to unexpected UNKNOWN_SERVER_ERROR. Retrying. (kafka.cluster.Partition)
[2022-02-28 04:01:20,217] ERROR [broker-1-to-controller] Uncaught error in request completion: (org.apache.kafka.clients.NetworkClient)
java.lang.IllegalStateException: Failed to enqueue `AlterIsr` request with state LeaderAndIsr(leader=1, leaderEpoch=2728, isr=List(1, 2), zkVersion=4719) for partition __consumer_offsets-4
at kafka.cluster.Partition.sendAlterIsrRequest(Partition.scala:1403)
at kafka.cluster.Partition.$anonfun$handleAlterIsrResponse$1(Partition.scala:1438)
at kafka.cluster.Partition.handleAlterIsrResponse(Partition.scala:1417)
at kafka.cluster.Partition.$anonfun$sendAlterIsrRequest$1(Partition.scala:1398)
at kafka.cluster.Partition.$anonfun$sendAlterIsrRequest$1$adapted(Partition.scala:1398)
at kafka.server.AlterIsrManagerImpl.$anonfun$handleAlterIsrResponse$8(AlterIsrManager.scala:166)
at kafka.server.AlterIsrManagerImpl.$anonfun$handleAlterIsrResponse$8$adapted(AlterIsrManager.scala:163)
at scala.collection.immutable.List.foreach(List.scala:333)
at kafka.server.AlterIsrManagerImpl.handleAlterIsrResponse(AlterIsrManager.scala:163)
at kafka.server.AlterIsrManagerImpl.responseHandler$1(AlterIsrManager.scala:94)
at kafka.server.AlterIsrManagerImpl.$anonfun$sendRequest$2(AlterIsrManager.scala:104)
at kafka.server.BrokerToControllerRequestThread.handleResponse(BrokerToControllerChannelManagerImpl.scala:175)
at kafka.server.BrokerToControllerRequestThread.$anonfun$generateRequests$1(BrokerToControllerChannelManagerImpl.scala:158)
at org.apache.kafka.clients.ClientResponse.onComplete(ClientResponse.java:109)
at org.apache.kafka.clients.NetworkClient.completeResponses(NetworkClient.java:586)
at org.apache.kafka.clients.NetworkClient.poll(NetworkClient.java:578)
at kafka.common.InterBrokerSendThread.doWork(InterBrokerSendThread.scala:71)
at kafka.server.BrokerToControllerRequestThread.doWork(BrokerToControllerChannelManagerImpl.scala:183)
at kafka.utils.ShutdownableThread.run(ShutdownableThread.scala:96)
 
under replication count goes to zero after the controller broker is restarted again. but this require manual intervention.
Expectation is that when broker reconnect with zookeeper cluster should come back to stable state with under replication count as zero by itself without any manual intervention.

Issue Links

• duplicates: KAFKA-13483 Stale Missing ISR on a partition after a zookeeper timeout