Kafka Request Handler deadlocks and brings down the cluster.

We have been experiencing a deadlock that happens on a consistent server within our cluster. This happens multiple times a week currently. It first started happening when we upgraded to 0.11.0.0. Sadly 0.11.0.1 failed to resolve the issue.

Sequence of events:

At a seemingly random time broker 125 goes into a deadlock. As soon as it is deadlocked it will remove all the ISR's for any partition is its the leader for.

[2017-10-10 00:06:10,061] INFO Partition [XXXXXXXXXX,24] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,073] INFO Partition [XXXXXXXXXX,974] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,079] INFO Partition [XXXXXXXXXX,64] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,081] INFO Partition [XXXXXXXXXX,21] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,084] INFO Partition [XXXXXXXXXX,12] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,085] INFO Partition [XXXXXXXXXX,61] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,086] INFO Partition [XXXXXXXXXX,53] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,088] INFO Partition [XXXXXXXXXX,27] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,090] INFO Partition [XXXXXXXXXX,182] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
[2017-10-10 00:06:10,091] INFO Partition [XXXXXXXXXX,16] on broker 125: Shrinking ISR from 117,125 to 125 (kafka.cluster.Partition)
....

The other nodes fail to connect to the node 125

[2017-10-10 00:08:42,318] WARN [ReplicaFetcherThread-0-125]: Error in fetch to broker 125, request (type=FetchRequest, replicaId=101, maxWait=500, minBytes=1, maxBytes=10485760, fetchData=

{XXXXXXXXXX-94=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-22=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-58=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-11=(offset=78932482, logStartOffset=50881481, maxBytes=1048576), XXXXXXXXXX-55=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-19=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-91=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-5=(offset=903857106, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-80=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-88=(offset=0, logStartOffset=0, maxBytes=1048576), XXXXXXXXXX-34=(offset=308, logStartOffset=308, maxBytes=1048576), XXXXXXXXXX-7=(offset=369990, logStartOffset=369990, maxBytes=1048576), XXXXXXXXXX-0=(offset=57965795, logStartOffset=0, maxBytes=1048576)}

) (kafka.server.ReplicaFetcherThread)
java.io.IOException: Connection to 125 was disconnected before the response was read
at org.apache.kafka.clients.NetworkClientUtils.sendAndReceive(NetworkClientUtils.java:93)
at kafka.server.ReplicaFetcherBlockingSend.sendRequest(ReplicaFetcherBlockingSend.scala:93)
at kafka.server.ReplicaFetcherThread.fetch(ReplicaFetcherThread.scala:207)
at kafka.server.ReplicaFetcherThread.fetch(ReplicaFetcherThread.scala:42)
at kafka.server.AbstractFetcherThread.processFetchRequest(AbstractFetcherThread.scala:151)
at kafka.server.AbstractFetcherThread.doWork(AbstractFetcherThread.scala:112)
at kafka.utils.ShutdownableThread.run(ShutdownableThread.scala:64)

As node 125 removed all the ISRs as it was locking up, a failover for any partition without an unclean leader election is not possible. This breaks any partition that this node was leader for. As we spread all topics across all servers this essentials brings down the entire cluster.

Recovery:

Unforuntately with broker 125 in its deadlocked state a clean shutdown does not work. A kill -9 is necessary. After an unclean shutdown the indexes must be rebuilt and start up time is around 10 minutes. After node 125 finally starts the cluster recovers.

A thread dump on 125 shows:

Found one Java-level deadlock:
=============================
"executor-Produce":
waiting to lock monitor 0x00007f01e417ac48 (object 0x000000068d10a1a0, a kafka.coordinator.group.GroupMetadata),
which is held by "kafka-request-handler-2"
"kafka-request-handler-2":
waiting to lock monitor 0x00007f0208f5e198 (object 0x000000068d2a45f8, a kafka.coordinator.group.GroupMetadata),
which is held by "kafka-request-handler-1"
"kafka-request-handler-1":
waiting to lock monitor 0x00007f01e417ac48 (object 0x000000068d10a1a0, a kafka.coordinator.group.GroupMetadata),
which is held by "kafka-request-handler-2"

Java stack information for the threads listed above:
===================================================
"executor-Produce":
at kafka.coordinator.group.GroupMetadataManager.putCacheCallback$2(GroupMetadataManager.scala:312)

• waiting to lock <0x000000068d10a1a0> (a kafka.coordinator.group.GroupMetadata)
  at kafka.coordinator.group.GroupMetadataManager.$anonfun$storeOffsets$10(GroupMetadataManager.scala:381)
  at kafka.coordinator.group.GroupMetadataManager.$anonfun$storeOffsets$10$adapted(GroupMetadataManager.scala:381)
  at kafka.coordinator.group.GroupMetadataManager$$Lambda$990/350831638.apply(Unknown Source)
  at kafka.server.DelayedProduce.onComplete(DelayedProduce.scala:131)
  at kafka.server.DelayedOperation.forceComplete(DelayedOperation.scala:66)
  at kafka.server.DelayedOperation.run(DelayedOperation.scala:112)
  at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
  at java.util.concurrent.FutureTask.run(FutureTask.java:266)
  at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
  at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
  at java.lang.Thread.run(Thread.java:748)
  "kafka-request-handler-2":
  at kafka.server.DelayedProduce.safeTryComplete(DelayedProduce.scala:75)
• waiting to lock <0x000000068d2a45f8> (a kafka.coordinator.group.GroupMetadata)
  at kafka.server.DelayedOperationPurgatory$Watchers.tryCompleteWatched(DelayedOperation.scala:338)
  at kafka.server.DelayedOperationPurgatory.checkAndComplete(DelayedOperation.scala:244)
  at kafka.server.ReplicaManager.tryCompleteDelayedProduce(ReplicaManager.scala:250)
  at kafka.cluster.Partition.tryCompleteDelayedRequests(Partition.scala:418)
  at kafka.cluster.Partition.appendRecordsToLeader(Partition.scala:500)
  at kafka.server.ReplicaManager.$anonfun$appendToLocalLog$2(ReplicaManager.scala:546)
  at kafka.server.ReplicaManager$$Lambda$887/1936159787.apply(Unknown Source)
  at scala.collection.TraversableLike.$anonfun$map$1(TraversableLike.scala:234)
  at scala.collection.TraversableLike$$Lambda$14/1859039536.apply(Unknown Source)
  at scala.collection.immutable.Map$Map1.foreach(Map.scala:120)
  at scala.collection.TraversableLike.map(TraversableLike.scala:234)
  at scala.collection.TraversableLike.map$(TraversableLike.scala:227)
  at scala.collection.AbstractTraversable.map(Traversable.scala:104)
  at kafka.server.ReplicaManager.appendToLocalLog(ReplicaManager.scala:532)
  at kafka.server.ReplicaManager.appendRecords(ReplicaManager.scala:374)
  at kafka.coordinator.group.GroupMetadataManager.appendForGroup(GroupMetadataManager.scala:246)
  at kafka.coordinator.group.GroupMetadataManager.storeOffsets(GroupMetadataManager.scala:381)
  at kafka.coordinator.group.GroupCoordinator.doCommitOffsets(GroupCoordinator.scala:465)
• locked <0x000000068d10a1a0> (a kafka.coordinator.group.GroupMetadata)
  at kafka.coordinator.group.GroupCoordinator.handleCommitOffsets(GroupCoordinator.scala:429)
  at kafka.server.KafkaApis.handleOffsetCommitRequest(KafkaApis.scala:361)
  at kafka.server.KafkaApis.handle(KafkaApis.scala:105)
  at kafka.server.KafkaRequestHandler.run(KafkaRequestHandler.scala:66)
  at java.lang.Thread.run(Thread.java:748)
  "kafka-request-handler-1":
  at kafka.server.DelayedProduce.safeTryComplete(DelayedProduce.scala:75)
• waiting to lock <0x000000068d10a1a0> (a kafka.coordinator.group.GroupMetadata)
  at kafka.server.DelayedOperationPurgatory$Watchers.tryCompleteWatched(DelayedOperation.scala:338)
  at kafka.server.DelayedOperationPurgatory.checkAndComplete(DelayedOperation.scala:244)
  at kafka.server.ReplicaManager.tryCompleteDelayedProduce(ReplicaManager.scala:250)
  at kafka.cluster.Partition.tryCompleteDelayedRequests(Partition.scala:418)
  at kafka.cluster.Partition.appendRecordsToLeader(Partition.scala:500)
  at kafka.server.ReplicaManager.$anonfun$appendToLocalLog$2(ReplicaManager.scala:546)
  at kafka.server.ReplicaManager$$Lambda$887/1936159787.apply(Unknown Source)
  at scala.collection.TraversableLike.$anonfun$map$1(TraversableLike.scala:234)
  at scala.collection.TraversableLike$$Lambda$14/1859039536.apply(Unknown Source)
  at scala.collection.immutable.Map$Map1.foreach(Map.scala:120)
  at scala.collection.TraversableLike.map(TraversableLike.scala:234)
  at scala.collection.TraversableLike.map$(TraversableLike.scala:227)
  at scala.collection.AbstractTraversable.map(Traversable.scala:104)
  at kafka.server.ReplicaManager.appendToLocalLog(ReplicaManager.scala:532)
  at kafka.server.ReplicaManager.appendRecords(ReplicaManager.scala:374)
  at kafka.coordinator.group.GroupMetadataManager.appendForGroup(GroupMetadataManager.scala:246)
  at kafka.coordinator.group.GroupMetadataManager.storeOffsets(GroupMetadataManager.scala:381)
  at kafka.coordinator.group.GroupCoordinator.doCommitOffsets(GroupCoordinator.scala:465)
• locked <0x000000068d2a45f8> (a kafka.coordinator.group.GroupMetadata)
  at kafka.coordinator.group.GroupCoordinator.handleCommitOffsets(GroupCoordinator.scala:429)
  at kafka.server.KafkaApis.handleOffsetCommitRequest(KafkaApis.scala:361)
  at kafka.server.KafkaApis.handle(KafkaApis.scala:105)
  at kafka.server.KafkaRequestHandler.run(KafkaRequestHandler.scala:66)
  at java.lang.Thread.run(Thread.java:748)

Full thread dump attached.