Dhruvil Shah created KAFKA-9307:
-----------------------------------
Summary: Transaction coordinator could be left in unknown state
after ZK session timeout
Key: KAFKA-9307
URL: https://issues.apache.org/jira/browse/KAFKA-9307
Project: Kafka
Issue Type: Bug
Components: core
Reporter: Dhruvil Shah
We observed a case where the transaction coordinator could not load transaction
state from __transaction-state topic partition. Clients would continue seeing
COORDINATOR_LOAD_IN_PROGRESS exceptions until the broker hosting the
coordinator is restarted.
This is the sequence of events that leads to the issue:
* The broker is the leader of one (or more) transaction state topic partitions.
* The broker loses its ZK session due to a network issue.
* Broker reestablishes session with ZK, though there are still transient
network issues.
* Broker is made follower of the transaction state topic partition it was
leading earlier.
* During the become-follower transition, the broker loses its ZK session again.
* The become-follower transition for this broker fails in-between, leaving us
in a partial leader / partial follower state for the transaction topic. This
meant that we could not unload the transaction metadata. However, the broker
successfully caches the leader epoch of associated with the LeaderAndIsrRequest.
```
[2019-12-12 03:08:17,864] ERROR [KafkaApi-3] Error when handling request:
clientId=2, correlationId=1, api=LEADER_AND_ISR, ...
\{topic=__transaction_state,partition_states=[{...
{partition=41,controller_epoch=16,leader=4,leader_epoch=112,isr=[2,4,1],zk_version=208,replicas=[3,4,2,1],is_new=false}
... org.apache.zookeeper.KeeperException$SessionExpiredException:
KeeperErrorCode = Session expired for /brokers/topics/__transaction_state at
org.apache.zookeeper.KeeperException.create(KeeperException.java:130) at
org.apache.zookeeper.KeeperException.create(KeeperException.java:54) at
kafka.zookeeper.AsyncResponse.resultException(ZooKeeperClient.scala:537) at
kafka.zk.KafkaZkClient$$anonfun$getReplicaAssignmentForTopics$1.apply(KafkaZkClient.scala:579)
at
kafka.zk.KafkaZkClient$$anonfun$getReplicaAssignmentForTopics$1.apply(KafkaZkClient.scala:574)
at
scala.collection.TraversableLike$$anonfun$flatMap$1.apply(TraversableLike.scala:241)
at
scala.collection.TraversableLike$$anonfun$flatMap$1.apply(TraversableLike.scala:241)
at
scala.collection.mutable.ResizableArray$class.foreach(ResizableArray.scala:59)
at scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:48) at
scala.collection.TraversableLike$class.flatMap(TraversableLike.scala:241) at
scala.collection.AbstractTraversable.flatMap(Traversable.scala:104) at
kafka.zk.KafkaZkClient.getReplicaAssignmentForTopics(KafkaZkClient.scala:574)
at kafka.zk.KafkaZkClient.getTopicPartitionCount(KafkaZkClient.scala:624) at
kafka.coordinator.transaction.TransactionStateManager.getTransactionTopicPartitionCount(TransactionStateManager.scala:279)
at
kafka.coordinator.transaction.TransactionStateManager.validateTransactionTopicPartitionCountIsStable(TransactionStateManager.scala:465)
at
kafka.coordinator.transaction.TransactionStateManager.removeTransactionsForTxnTopicPartition(TransactionStateManager.scala:434)
at
kafka.coordinator.transaction.TransactionCoordinator.handleTxnEmigration(TransactionCoordinator.scala:282)
at
kafka.server.KafkaApis$$anonfun$kafka$server$KafkaApis$$onLeadershipChange$1$2.apply(KafkaApis.scala:190)
at
kafka.server.KafkaApis$$anonfun$kafka$server$KafkaApis$$onLeadershipChange$1$2.apply(KafkaApis.scala:186)
at scala.collection.mutable.HashSet.foreach(HashSet.scala:78) at
kafka.server.KafkaApis.kafka$server$KafkaApis$$onLeadershipChange$1(KafkaApis.scala:186)
at kafka.server.KafkaApis$$anonfun$2.apply(KafkaApis.scala:202) at
kafka.server.KafkaApis$$anonfun$2.apply(KafkaApis.scala:202) at
kafka.server.ReplicaManager.becomeLeaderOrFollower(ReplicaManager.scala:1153)
at kafka.server.KafkaApis.handleLeaderAndIsrRequest(KafkaApis.scala:202)
```
* Later, when the ZK session is finally established successfully, the broker
ignores the become-follower transition as the leader epoch was same as the one
it had cached. This prevented the transaction metadata from being unloaded.
* Because this partition was a partial follower, we had setup replica
fetchers. The partition continued to fetch from the leader until it was made
part of the ISR.
* Once it was part of the ISR, preferred leader election kicked in and elected
this broker as the leader.
* When processing the become-leader transition, the operation failed as we
already had transaction metadata loaded at a previous epoch. This meant that
this partition was left in the "loading" state and we thus returned
COORDINATOR_LOAD_IN_PROGRESS errors.
* Broker restart fixed this partial in-memory state and we were able to resume
processing for transactions.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
