Kafka Partition重分配流程简析
节日快乐~
今天是属于广大程序员的节日，祝自己快乐hhhhhh
随着业务量的急速膨胀和又一年双11的到来，我们会对现有的Kafka集群进行扩容，以应对更大的流量和业务尖峰。当然，扩容之后的新Kafka Broker默认是不会有任何Topic和Partition的，需要手动利用分区重分配命令kafka-reassign-partitions将现有的Partition/Replica平衡到新的Broker上去。那么Kafka具体是如何执行重分配流程的呢？本文就来简单解读一下。
生成、提交重分配方案
我们知道，使用kafka-reassign-partitions命令分为三步，一是根据指定的Topic生成JSON格式的重分配方案（--generate），二是将生成的方案提交执行（--execute），三是观察重分配的进度（--verify），它们分别对应kafka.admin.ReassignPartitionsCommand类中的generateAssignment()、executeAssignment()和verifyAssignment()方法。
generateAssignment()方法会调用AdminUtils#assignReplicasToBrokers()方法生成Replica分配方案。源码就不再读了，其原则简述如下：
- 将Replica尽量均匀地分配到各个Broker上去；
- 一个Partition的所有Replica必须位于不同的Broker上；
- 如果Broker有机架感知（rack aware）的信息，将Partition的Replica尽量分配到不同的机架。
executeReassignment()方法调用了reassignPartitions()方法，其源码如下。
def reassignPartitions(throttle: Throttle = NoThrottle, timeoutMs: Long = 10000L): Boolean = {
maybeThrottle(throttle)
try {
val validPartitions = proposedPartitionAssignment.filter { case (p, _) => validatePartition(zkUtils, p.topic, p.partition) }
if (validPartitions.isEmpty) false
else {
if (proposedReplicaAssignment.nonEmpty) {
val adminClient = adminClientOpt.getOrElse(
throw new AdminCommandFailedException("bootstrap-server needs to be provided in order to reassign replica to the specified log directory"))
val alterReplicaDirResult = adminClient.alterReplicaLogDirs(
proposedReplicaAssignment.asJava, new AlterReplicaLogDirsOptions().timeoutMs(timeoutMs.toInt))
alterReplicaDirResult.values().asScala.foreach { case (replica, future) => {
try {
future.get()
throw new AdminCommandFailedException(s"Partition ${replica.topic()}-${replica.partition()} already exists on broker ${replica.brokerId()}." +
s" Reassign replica to another log directory on the same broker is currently not supported.")
} catch {
case t: ExecutionException =>
t.getCause match {
case e: ReplicaNotAvailableException => // It is OK if the replica is not available
case e: Throwable => throw e
}
}
}}
}
val jsonReassignmentData = ZkUtils.formatAsReassignmentJson(validPartitions)
zkUtils.createPersistentPath(ZkUtils.ReassignPartitionsPath, jsonReassignmentData)
true
}
} catch {
// ......
}
}
在进行必要的Partition校验之后，创建ZK持久节点/admin/reassign_partitions，并将JSON格式的重分配方案写进去。如果该节点存在，就表示已经在进行重分配，不能再启动新的重分配流程（相关的判断在executeReassignment()方法中）。
监听并处理重分配事件
在之前讲解Kafka Controller时，笔者提到Controller会注册多个ZK监听器，将监听到的事件投递到内部的事件队列，并由事件处理线程负责处理。监听ZK中/admin/reassign_partitions节点的监听器为PartitionReassignmentListener，并产生PartitionReassignment事件，处理逻辑如下。
case object PartitionReassignment extends ControllerEvent {
def state = ControllerState.PartitionReassignment
override def process(): Unit = {
if (!isActive) return
val partitionReassignment = zkUtils.getPartitionsBeingReassigned()
val partitionsToBeReassigned = partitionReassignment.filterNot(p => controllerContext.partitionsBeingReassigned.contains(p._1))
partitionsToBeReassigned.foreach { case (partition, context) =>
if(topicDeletionManager.isTopicQueuedUpForDeletion(partition.topic)) {
error(s"Skipping reassignment of partition $partition since it is currently being deleted")
removePartitionFromReassignedPartitions(partition)
} else {
initiateReassignReplicasForTopicPartition(partition, context)
}
}
}
}
该方法先取得需要重分配的Partition列表，然后从中剔除掉那些已经被标记为删除的Topic所属的Partition，再调用initiateReassignReplicasForTopicPartition()方法：
def initiateReassignReplicasForTopicPartition(topicAndPartition: TopicAndPartition,
reassignedPartitionContext: ReassignedPartitionsContext) {
val newReplicas = reassignedPartitionContext.newReplicas
val topic = topicAndPartition.topic
val partition = topicAndPartition.partition
try {
val assignedReplicasOpt = controllerContext.partitionReplicaAssignment.get(topicAndPartition)
assignedReplicasOpt match {
case Some(assignedReplicas) =>
if (assignedReplicas == newReplicas) {
throw new KafkaException("Partition %s to be reassigned is already assigned to replicas".format(topicAndPartition) +
" %s. Ignoring request for partition reassignment".format(newReplicas.mkString(",")))
} else {
info("Handling reassignment of partition %s to new replicas %s".format(topicAndPartition, newReplicas.mkString(",")))
// first register ISR change listener
watchIsrChangesForReassignedPartition(topic, partition, reassignedPartitionContext)
controllerContext.partitionsBeingReassigned.put(topicAndPartition, reassignedPartitionContext)
// mark topic ineligible for deletion for the partitions being reassigned
topicDeletionManager.markTopicIneligibleForDeletion(Set(topic))
onPartitionReassignment(topicAndPartition, reassignedPartitionContext)
}
case None => throw new KafkaException("Attempt to reassign partition %s that doesn't exist"
.format(topicAndPartition))
}
} catch {
case e: Throwable => error("Error completing reassignment of partition %s".format(topicAndPartition), e)
// remove the partition from the admin path to unblock the admin client
removePartitionFromReassignedPartitions(topicAndPartition)
}
}
该方法的执行逻辑如下：
- 判断Partition的原有Replica是否与即将重分配的新Replica相同，如果相同则抛出异常；
- 注册即将被重分配的Partition的ISR变化监听器；
- 把即将被重分配的Partition/Replica记录在Controller上下文中的partitionsBeingReassigned集合中；
- 把即将被重分配的Topic标记为不可删除；
- 调用onPartitionReassignment()方法真正触发重分配流程。
执行重分配流程
onPartitionReassignment()方法的代码如下。
def onPartitionReassignment(topicAndPartition: TopicAndPartition, reassignedPartitionContext: ReassignedPartitionsContext) {
val reassignedReplicas = reassignedPartitionContext.newReplicas
if (!areReplicasInIsr(topicAndPartition.topic, topicAndPartition.partition, reassignedReplicas)) {
info("New replicas %s for partition %s being ".format(reassignedReplicas.mkString(","), topicAndPartition) +
"reassigned not yet caught up with the leader")
val newReplicasNotInOldReplicaList = reassignedReplicas.toSet -- controllerContext.partitionReplicaAssignment(topicAndPartition).toSet
val newAndOldReplicas = (reassignedPartitionContext.newReplicas ++ controllerContext.partitionReplicaAssignment(topicAndPartition)).toSet
//1. Update AR in ZK with OAR + RAR.
updateAssignedReplicasForPartition(topicAndPartition, newAndOldReplicas.toSeq)
//2. Send LeaderAndIsr request to every replica in OAR + RAR (with AR as OAR + RAR).
updateLeaderEpochAndSendRequest(topicAndPartition, controllerContext.partitionReplicaAssignment(topicAndPartition),
newAndOldReplicas.toSeq)
//3. replicas in RAR - OAR -> NewReplica
startNewReplicasForReassignedPartition(topicAndPartition, reassignedPartitionContext, newReplicasNotInOldReplicaList)
info("Waiting for new replicas %s for partition %s being ".format(reassignedReplicas.mkString(","), topicAndPartition) +
"reassigned to catch up with the leader")
} else {
//4. Wait until all replicas in RAR are in sync with the leader.
val oldReplicas = controllerContext.partitionReplicaAssignment(topicAndPartition).toSet -- reassignedReplicas.toSet
//5. replicas in RAR -> OnlineReplica
reassignedReplicas.foreach { replica =>
replicaStateMachine.handleStateChanges(Set(PartitionAndReplica(topicAndPartition.topic, topicAndPartition.partition,
replica)), OnlineReplica)
}
//6. Set AR to RAR in memory.
//7. Send LeaderAndIsr request with a potential new leader (if current leader not in RAR) and
// a new AR (using RAR) and same isr to every broker in RAR
moveReassignedPartitionLeaderIfRequired(topicAndPartition, reassignedPartitionContext)
//8. replicas in OAR - RAR -> Offline (force those replicas out of isr)
//9. replicas in OAR - RAR -> NonExistentReplica (force those replicas to be deleted)
stopOldReplicasOfReassignedPartition(topicAndPartition, reassignedPartitionContext, oldReplicas)
//10. Update AR in ZK with RAR.
updateAssignedReplicasForPartition(topicAndPartition, reassignedReplicas)
//11. Update the /admin/reassign_partitions path in ZK to remove this partition.
removePartitionFromReassignedPartitions(topicAndPartition)
info("Removed partition %s from the list of reassigned partitions in zookeeper".format(topicAndPartition))
controllerContext.partitionsBeingReassigned.remove(topicAndPartition)
//12. After electing leader, the replicas and isr information changes, so resend the update metadata request to every broker
sendUpdateMetadataRequest(controllerContext.liveOrShuttingDownBrokerIds.toSeq, Set(topicAndPartition))
// signal delete topic thread if reassignment for some partitions belonging to topics being deleted just completed
topicDeletionManager.resumeDeletionForTopics(Set(topicAndPartition.topic))
}
}
官方JavaDoc比较详细，给出了3个方便解释流程的定义，列举如下：
- RAR（Re-assigned replicas）：重分配的Replica集合，记为reassignedReplicas；
- OAR（Original assigned replicas）：重分配之前的原始Replica集合，通过controllerContext.partitionReplicaAssignment()方法取得；
- AR（Assigned replicas）：当前的Replica集合，随着重分配的进行不断变化。
根据上文的代码和注释，我们可以很容易地梳理出重分配的具体流程：
(0) 检查RAR是否都已经在Partition的ISR集合中（即是否已经同步），若否，则计算RAR与OAR的差集，也就是需要被创建或者重分配的Replica集合；
(1) 计算RAR和OAR的并集，即所有Replica的集合，并将ZK中的AR更新；
(2) 增加Partition的Leader纪元值，并向AR中的所有Replica所在的Broker发送LeaderAndIsrRequest；
(3) 更新RAR与OAR的差集中Replica的状态为NewReplica，以触发这些Replica的创建或同步；
(4) 计算OAR和RAR的差集，即重分配过程中需要被下线的Replica集合；
(5) 等待RAR都已经在Partition的ISR集合中，将RAR中Replica的状态设置为OnlineReplica，表示同步完成；
(6) 将迁移现场的AR更新为RAR；
(7) 检查Partition的Leader是否在RAR中，如果没有，则触发新的Leader选举。然后增加Partition的Leader纪元值，发送LeaderAndIsrRequest更新Leader的结果；
(8~9) 将OAR和RAR的差集中的Replica状态设为Offline->NonExistentReplica，这些Replica后续将被删除；
(10) 将ZK中的AR集合更新为RAR；
(11) 一个Partition重分配完成，更新/admin/reassign_partitions节点中的执行计划，删掉完成的Partition；
(12) 发送UpdateMetadataRequest给所有Broker，刷新元数据缓存；
(13) 如果有一个Topic已经重分配完成并且将被删除，就将它从不可删除的Topic集合中移除。
The End
最后一个小问题：Partition重分配往往会涉及大量的数据交换，有可能会影响正常业务的运行，如何避免呢？ReassignPartitionsCommand也提供了throttle功能用于限流，在代码和帮助文档中都可以看到它，就不多讲了。当然，一旦启用了throttle，我们一定要定期进行verify操作，防止因为限流导致重分配的Replica一直追不上Leader的情况发生。
民那晚安晚安。
