[KAFKA-5758] Reassigning a topic's partitions can adversely impact other topics

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.10.1.1
• Fix Version/s: 1.0.0
• Component/s: core
• Labels: reliability
• Assignee: Ismael Juma
• Reporter: David van Geest
• Created: 21/Aug/17 17:11
• Updated: 17/Mar/18 09:13
• Resolved: 02/Oct/17 21:44

Description

We've noticed that reassigning a topic's partitions seems to adversely impact other topics. Specifically, followers for other topics fall out of the ISR.

While I'm not 100% sure about why this happens, the scenario seems to be as follows:

1. Reassignment is manually triggered on topic-partition X-Y, and broker A (which used to be a follower for X-Y) is no longer a follower.
2. Broker A makes `FetchRequest` including topic-partition X-Y to broker B, just after the reassignment.
3. Broker B can fulfill the `FetchRequest`, but while trying to do so it tries to record the position of "follower" A. This fails, because broker A is no longer a follower for X-Y (see exception below).
4. The entire `FetchRequest` request fails, and broker A's other followed topics start falling behind.
5. Depending on the length of the reassignment, this sequence repeats.

In step 3, we see exceptions like:

Error when handling request Name: FetchRequest; Version: 3; CorrelationId: 46781859; ClientId: ReplicaFetcherThread-0-1001; ReplicaId: 1006; MaxWait: 500 ms; MinBytes: 1 bytes; MaxBytes:10485760 bytes; RequestInfo: 

<LOTS OF PARTITIONS>

kafka.common.NotAssignedReplicaException: Leader 1001 failed to record follower 1006's position -1 since the replica is not recognized to be one of the assigned replicas 1001,1004,1005 for partition [topic_being_reassigned,5].
at kafka.cluster.Partition.updateReplicaLogReadResult(Partition.scala:249)
	at kafka.server.ReplicaManager$$anonfun$updateFollowerLogReadResults$2.apply(ReplicaManager.scala:923)
	at kafka.server.ReplicaManager$$anonfun$updateFollowerLogReadResults$2.apply(ReplicaManager.scala:920)
	at scala.collection.mutable.ResizableArray$class.foreach(ResizableArray.scala:59)
	at scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:48)
	at kafka.server.ReplicaManager.updateFollowerLogReadResults(ReplicaManager.scala:920)
	at kafka.server.ReplicaManager.fetchMessages(ReplicaManager.scala:481)
	at kafka.server.KafkaApis.handleFetchRequest(KafkaApis.scala:534)
	at kafka.server.KafkaApis.handle(KafkaApis.scala:79)
	at kafka.server.KafkaRequestHandler.run(KafkaRequestHandler.scala:60)
	at java.lang.Thread.run(Thread.java:745)

Does my assessment make sense? If so, this behaviour seems problematic. A few changes that might improve matters (assuming I'm on the right track):

1. `FetchRequest` should be able to return partial results
2. The broker fulfilling the `FetchRequest` could ignore the `NotAssignedReplicaException`, and return results without recording the not-any-longer-follower position.

This behaviour was experienced with 0.10.1.1, although looking at the changelogs and the code in question, I don't see any reason why it would have changed in later versions.

Am very interested to have some discussion on this. Thanks!

Issue Links

• relates to: KAFKA-4027 Leader for a cetain partition unavailable forever
• relates to: KAFKA-6613 The controller shouldn't stop partition reassignment after an exception is being thrown
• links to: GitHub Pull Request #3954