KAFKA-6630: Speed up the processing of TopicDeletionStopReplicaResponseReceived events on the controller#4668
Conversation
|
@onurkaraman Can you please take a look at this patch? Thanks. |
|
@onurkaraman @junrao Can you please take a look at this patch? Thanks! |
| .getOrElse(topicPartition.partition, Seq.empty) | ||
| } | ||
|
|
||
| def clearPartitionReplicaAssignment() = { |
There was a problem hiding this comment.
To be consistent, for methods with no return value, we want to specify Unit =. Also, should we make this method more general to clear other fields such as allTopics, partitionsBeingReassigned and replicasOnOfflineDirs, instead of letting the caller do that?
There was a problem hiding this comment.
Added the return type Unit and converted the clearing method to clear other topic states.
| .put(topicPartition.partition, newReplicas) | ||
| } | ||
|
|
||
| def partitionReplicaAssignmentForTopic(topic : String) : mutable.Map[TopicPartition, Seq[Int]] = { |
There was a problem hiding this comment.
It's probably better not to return a mutable map.
| } | ||
| } | ||
|
|
||
| def removePartitionReplicaAssignmentForTopic(topic : String) = { |
There was a problem hiding this comment.
It seems that this is the same method as removeTopic? The latter seems more complete.
There was a problem hiding this comment.
Removed this method and replaced it with removeTopic instead.
| def removeTopic(topic: String) = { | ||
| partitionLeadershipInfo = partitionLeadershipInfo.filter { case (topicPartition, _) => topicPartition.topic != topic } | ||
| partitionReplicaAssignment = partitionReplicaAssignment.filter { case (topicPartition, _) => topicPartition.topic != topic } | ||
| partitionReplicaAssignmentUnderlying.remove(topic) |
There was a problem hiding this comment.
Should we also update replicasOnOfflineDirs accordingly?
There was a problem hiding this comment.
I think we should update replicasOnOfflineDirs accordingly. However I don't want to simply iterate through the current replicasOnOfflineDirs to filter out the topic since doing that will be very slow in a large cluster.
Since clearing the replicasOnOfflineDirs is an orthogonal change, I'll think more about whether it's important and if so how to do it. I feel a separate PR is probably better. Comment?
| partitionReplicaAssignmentUnderlying = mutable.Map.empty | ||
| } | ||
|
|
||
| def updatePartitionReplicaAssignment(topicPartition: TopicPartition, newReplicas : Seq[Int]) = { |
There was a problem hiding this comment.
To be consistent, we want to remove the space before :.
| }.keySet.map(_.topic) | ||
| val topicsWithOfflineReplicas = controllerContext.allTopics.filter { topic => { | ||
| val replicasForTopic = controllerContext.replicasForTopic(topic) | ||
| replicasForTopic.exists(r => !controllerContext.isReplicaOnline(r.replica, new TopicPartition(topic, r.partition))) |
There was a problem hiding this comment.
It seems that we could just use r.topicPartition instead of creating a new TopicPartition.
| 0 | ||
| } else { | ||
| controllerContext.partitionReplicaAssignment.count { case (topicPartition, replicas) => | ||
| controllerContext.allPartitions.count { topicAndPartition => |
There was a problem hiding this comment.
topicAndPartition => topicPartition
| val partitionReplicaAssignment = zkClient.getReplicaAssignmentForTopics(immutable.Set(topic)) | ||
| val partitionsToBeAdded = partitionReplicaAssignment.filter(p => | ||
| !controllerContext.partitionReplicaAssignment.contains(p._1)) | ||
| controllerContext.partitionReplicaAssignment(p._1).isEmpty) |
There was a problem hiding this comment.
Perhaps we can change p to case (topicPartion, _) and use topicPartition instead of p._1?
| if (partitionsToBeAdded.nonEmpty) { | ||
| info(s"New partitions to be added $partitionsToBeAdded") | ||
| controllerContext.partitionReplicaAssignment ++= partitionsToBeAdded | ||
| partitionsToBeAdded.foreach { case (topicAndPartition, assignedReplicas) => |
There was a problem hiding this comment.
topicAndPartition => topicPartition ?
|
@ijuma Yes, sorry about the delay. I'll update this PR soon. |
|
@junrao Can you please take another look? Thanks! |
| .getOrElse(topicPartition.partition, Seq.empty) | ||
| } | ||
|
|
||
| def clearTopicsState(): Unit = { |
| private var liveBrokersUnderlying: Set[Broker] = Set.empty | ||
| private var liveBrokerIdsUnderlying: Set[Int] = Set.empty | ||
|
|
||
| def partitionReplicaAssignment(topicPartition: TopicPartition) : Seq[Int] = { |
There was a problem hiding this comment.
Could we remove the space before : Seq[Int] and some other places?
3 participants
This patch tries to speed up the inefficient functions identified in Kafka-6630 by grouping partitions in the ControllerContext.partitionReplicaAssignment variable by topics. Hence trying to find all replicas for a topic won't need to go through all the replicas in the cluster.
Passed all tests using "gradle testAll"
Committer Checklist (excluded from commit message)