Apache Kafka源码分析 - PartitionStateMachine

startup

在onControllerFailover中被调用，

initializePartitionState

private def initializePartitionState() {
    for((topicPartition, replicaAssignment) <- controllerContext.partitionReplicaAssignment) { // 取出所有的partitions
      // check if leader and isr path exists for partition. If not, then it is in NEW state
      controllerContext.partitionLeadershipInfo.get(topicPartition) match {
        case Some(currentLeaderIsrAndEpoch) =>
          // else, check if the leader for partition is alive. If yes, it is in Online state, else it is in Offline state
          controllerContext.liveBrokerIds.contains(currentLeaderIsrAndEpoch.leaderAndIsr.leader) match {
            case true => // leader is alive
              partitionState.put(topicPartition, OnlinePartition)
            case false =>
              partitionState.put(topicPartition, OfflinePartition)
          }
        case None =>
          partitionState.put(topicPartition, NewPartition)
      }
    }
  }

这里注意offlinePartition和newPartition的区别，

如果controllerContext.partitionLeadershipInfo中没有这个partition的leader信息，那么说明是newPartition

如果有leader，但leader所在broker不是alive的，那么就是offlinePartition

当然，如果leader所在broker是alive的，那么就是onlinePartition

 

triggerOnlinePartitionStateChange

试图将所有offline和new partition的状态变成online

def triggerOnlinePartitionStateChange() {
    try {
      brokerRequestBatch.newBatch()
      // try to move all partitions in NewPartition or OfflinePartition state to OnlinePartition state except partitions
      // that belong to topics to be deleted
      for((topicAndPartition, partitionState) <- partitionState
          if(!controller.deleteTopicManager.isTopicQueuedUpForDeletion(topicAndPartition.topic))) {
        if(partitionState.equals(OfflinePartition) || partitionState.equals(NewPartition))
          handleStateChange(topicAndPartition.topic, topicAndPartition.partition, OnlinePartition, controller.offlinePartitionSelector,
                            (new CallbackBuilder).build)
      }
      brokerRequestBatch.sendRequestsToBrokers(controller.epoch, controllerContext.correlationId.getAndIncrement)
    } catch {
      case e: Throwable => error("Error while moving some partitions to the online state", e)
      // TODO: It is not enough to bail out and log an error, it is important to trigger leader election for those partitions
    }
  }

这里看到，brokerRequestBatch，这个经常出现，ControllerBrokerRequestBatch

这个类封装了leaderAndIsrRequestMap，stopReplicaRequestMap，updateMetadataRequestMap

用来记录和cache，handleStateChange中产生的这些request

最终用sendRequestsToBrokers，将这些requests，批量的发出去

handleStateChange的逻辑后面单独看，这里看看controller.offlinePartitionSelector，这个selector实现如何为一个newPartition或offlinePartition选一个leader

代码挺长的，注释讲的挺清楚的，就不贴代码了

首先如果ISR里面有活的broker，那没有好说的，直接用它作为新的leader

如果没有，这里需要看一下是否容忍unclean leader election，其实就是是否可以容忍丢数据，如果可以

那么就看看这AR里面有没有活的broker，如果有就以它为leader，但这个既然不在ISR里面，说明这个replica是不同步的，所以一定有data loss

如果AR里面也没有活的broker，那只能是elect失败了

/**
 * Select the new leader, new isr and receiving replicas (for the LeaderAndIsrRequest):
 * 1. If at least one broker from the isr is alive, it picks a broker from the live isr as the new leader and the live
 *    isr as the new isr.
 * 2. Else, if unclean leader election for the topic is disabled, it throws a NoReplicaOnlineException.
 * 3. Else, it picks some alive broker from the assigned replica list as the new leader and the new isr.
 * 4. If no broker in the assigned replica list is alive, it throws a NoReplicaOnlineException
 * Replicas to receive LeaderAndIsr request = live assigned replicas
 * Once the leader is successfully registered in zookeeper, it updates the allLeaders cache
 */

registerListeners

在onControllerFailover中被调用，

这里负责注册一下listener到zk，deleteTopicListener先不管

先看看TopicChangeListener，当topics发生变化时，我们做什么处理？

 

registerTopicChangeListener

private def registerTopicChangeListener() = {
    zkClient.subscribeChildChanges(ZkUtils.BrokerTopicsPath, topicChangeListener) //"/brokers/topics"
  }

Listen这个目录， /brokers/topics，如果发生变化，触发topicChangeListener

 

TopicChangeListener

  /**
   * This is the zookeeper listener that triggers all the state transitions for a partition
   */
  class TopicChangeListener extends IZkChildListener with Logging {
    this.logIdent = "[TopicChangeListener on Controller " + controller.config.brokerId + "]: "

    @throws(classOf[Exception])
    def handleChildChange(parentPath : String, children : java.util.List[String]) {
      inLock(controllerContext.controllerLock) {
        if (hasStarted.get) {
          try {
            val currentChildren = {
              import JavaConversions._
              debug("Topic change listener fired for path %s with children %s".format(parentPath, children.mkString(",")))
              (children: Buffer[String]).toSet
            }
            val newTopics = currentChildren -- controllerContext.allTopics //context里面没记录，但zk有的，就是新topic
            val deletedTopics = controllerContext.allTopics -- currentChildren //反之，就被删除的topic
            controllerContext.allTopics = currentChildren //更新context

            val addedPartitionReplicaAssignment = ZkUtils.getReplicaAssignmentForTopics(zkClient, newTopics.toSeq) //从zk取出新topic的assignment情况
              controllerContext.partitionReplicaAssignment = controllerContext.partitionReplicaAssignment.filter(p =>  //从context中的assignment情况中删掉deletedtopic的
              !deletedTopics.contains(p._1.topic))
            controllerContext.partitionReplicaAssignment.++=(addedPartitionReplicaAssignment) //把新的topic的assignment加入context
            info("New topics: [%s], deleted topics: [%s], new partition replica assignment [%s]".format(newTopics,
              deletedTopics, addedPartitionReplicaAssignment))
            if(newTopics.size > 0)
              controller.onNewTopicCreation(newTopics, addedPartitionReplicaAssignment.keySet.toSet) //最终调用KafkaController.onNewTopicCreation
          } catch {
            case e: Throwable => error("Error while handling new topic", e )
          }
        }
      }
    }
  }

 

onNewTopicCreation

def onNewTopicCreation(topics: Set[String], newPartitions: Set[TopicAndPartition]) {
    info("New topic creation callback for %s".format(newPartitions.mkString(",")))
    // subscribe to partition changes
    topics.foreach(topic => partitionStateMachine.registerPartitionChangeListener(topic)) //添加partition change listener
    onNewPartitionCreation(newPartitions) //partition和replica的状态变化
  }

def registerPartitionChangeListener(topic: String) = {
    addPartitionsListener.put(topic, new AddPartitionsListener(topic))
    zkClient.subscribeDataChanges(ZkUtils.getTopicPath(topic), addPartitionsListener(topic)) ///brokers/topics/topic-name
  }

 

AddPartitionsListener

和topic listener很想，就是从zk读出partition情况，和当前context里面的比较，找出新的partitions，调用

controller.onNewPartitionCreation(partitionsToBeAdded.keySet.toSet)

可见无论是TopicChangeListener还是AddPartitionsListener，最终都是调用到onNewPartitionCreation，毕竟topic是个逻辑概念

onNewPartitionCreation

def onNewPartitionCreation(newPartitions: Set[TopicAndPartition]) {
    info("New partition creation callback for %s".format(newPartitions.mkString(",")))
    partitionStateMachine.handleStateChanges(newPartitions, NewPartition)
    replicaStateMachine.handleStateChanges(controllerContext.replicasForPartition(newPartitions), NewReplica)
    partitionStateMachine.handleStateChanges(newPartitions, OnlinePartition, offlinePartitionSelector)
    replicaStateMachine.handleStateChanges(controllerContext.replicasForPartition(newPartitions), OnlineReplica)
  }

很简单，只是首先将所有新的partition和相应的replica的状态设为new，然后再设为online

 

handleStateChange

这是状态机的主逻辑，

private def handleStateChange(topic: String, partition: Int, targetState: PartitionState,
                                leaderSelector: PartitionLeaderSelector,
                                callbacks: Callbacks) {
    val topicAndPartition = TopicAndPartition(topic, partition)
    val currState = partitionState.getOrElseUpdate(topicAndPartition, NonExistentPartition) // 取得当前状态
    try {
      targetState match {
        case NewPartition =>
          // pre: partition did not exist before this
          assertValidPreviousStates(topicAndPartition, List(NonExistentPartition), NewPartition)
          assignReplicasToPartitions(topic, partition) // 从zk取得AR，并更新controllerContext.partitionReplicaAssignment
          partitionState.put(topicAndPartition, NewPartition)
          // post: partition has been assigned replicas
        case OnlinePartition =>
          assertValidPreviousStates(topicAndPartition, List(NewPartition, OnlinePartition, OfflinePartition), OnlinePartition)
          partitionState(topicAndPartition) match {
            case NewPartition =>
              // initialize leader and isr path for new partition
              initializeLeaderAndIsrForPartition(topicAndPartition)
            case OfflinePartition =>
              electLeaderForPartition(topic, partition, leaderSelector)
            case OnlinePartition => // invoked when the leader needs to be re-elected
              electLeaderForPartition(topic, partition, leaderSelector)
            case _ => // should never come here since illegal previous states are checked above
          }
          partitionState.put(topicAndPartition, OnlinePartition)
          val leader = controllerContext.partitionLeadershipInfo(topicAndPartition).leaderAndIsr.leader
          stateChangeLogger.trace("Controller %d epoch %d changed partition %s from %s to %s with leader %d"
                                    .format(controllerId, controller.epoch, topicAndPartition, currState, targetState, leader))
           // post: partition has a leader
        case OfflinePartition =>
          // pre: partition should be in New or Online state
          assertValidPreviousStates(topicAndPartition, List(NewPartition, OnlinePartition, OfflinePartition), OfflinePartition)
          // should be called when the leader for a partition is no longer alive
          stateChangeLogger.trace("Controller %d epoch %d changed partition %s state from %s to %s"
                                    .format(controllerId, controller.epoch, topicAndPartition, currState, targetState))
          partitionState.put(topicAndPartition, OfflinePartition)
          // post: partition has no alive leader
        case NonExistentPartition =>
          // pre: partition should be in Offline state
          assertValidPreviousStates(topicAndPartition, List(OfflinePartition), NonExistentPartition)
          stateChangeLogger.trace("Controller %d epoch %d changed partition %s state from %s to %s"
                                    .format(controllerId, controller.epoch, topicAndPartition, currState, targetState))
          partitionState.put(topicAndPartition, NonExistentPartition)
          // post: partition state is deleted from all brokers and zookeeper
      }
    } catch {
      case t: Throwable =>
        stateChangeLogger.error("Controller %d epoch %d initiated state change for partition %s from %s to %s failed"
          .format(controllerId, controller.epoch, topicAndPartition, currState, targetState), t)
    }
  }

可以看到，对于转变到OfflinePartition，NonExistentPartition，只是单纯的设置state

而转变到NewPartition，除了设置state，也就多了步初始化AR

只有转变到OnlinePartition的时候比较复杂些，

其中从NewPartition--》OnlinePartition，需要做些初始化的工作，所以调用initializeLeaderAndIsrForPartition

initializeLeaderAndIsrForPartition

NewPartition是在zk中，没有leaderAndISR path的，所以初始化需要创建path，创建后，就再也不能回到New的状态，只能到offline

其中逻辑除了创建zk path，就是进行leader elect，这里的elect逻辑是写死的，初始化的时候，一定是prefered selector，即选取live AR的head

/**
   * Invoked on the NewPartition->OnlinePartition state change. When a partition is in the New state, it does not have
   * a leader and isr path in zookeeper. Once the partition moves to the OnlinePartition state, it's leader and isr
   * path gets initialized and it never goes back to the NewPartition state. From here, it can only go to the
   * OfflinePartition state.
   * @param topicAndPartition   The topic/partition whose leader and isr path is to be initialized
   */
  private def initializeLeaderAndIsrForPartition(topicAndPartition: TopicAndPartition) {
    val replicaAssignment = controllerContext.partitionReplicaAssignment(topicAndPartition)
    val liveAssignedReplicas = replicaAssignment.filter(r => controllerContext.liveBrokerIds.contains(r)) // 找出AR中活着的replica
    liveAssignedReplicas.size match {
      case 0 => // 没有活的，那肯定无法转成online的
        val failMsg = ("encountered error during state change of partition %s from New to Online, assigned replicas are [%s], " +
                       "live brokers are [%s]. No assigned replica is alive.")
                         .format(topicAndPartition, replicaAssignment.mkString(","), controllerContext.liveBrokerIds)
        stateChangeLogger.error("Controller %d epoch %d ".format(controllerId, controller.epoch) + failMsg)
        throw new StateChangeFailedException(failMsg)
      case _ =>
        debug("Live assigned replicas for partition %s are: [%s]".format(topicAndPartition, liveAssignedReplicas))
        // make the first replica in the list of assigned replicas, the leader
        val leader = liveAssignedReplicas.head  // 取出第一个活的replica，作为leader replica
        val leaderIsrAndControllerEpoch = new LeaderIsrAndControllerEpoch(new LeaderAndIsr(leader, liveAssignedReplicas.toList), // 封装出LeaderIsrAndControllerEpoch
          controller.epoch)
        debug("Initializing leader and isr for partition %s to %s".format(topicAndPartition, leaderIsrAndControllerEpoch))
        try {
          ZkUtils.createPersistentPath(controllerContext.zkClient,  // 创建zk的LeaderAndIsrPath，关键的初始步骤
            ZkUtils.getTopicPartitionLeaderAndIsrPath(topicAndPartition.topic, topicAndPartition.partition),
            ZkUtils.leaderAndIsrZkData(leaderIsrAndControllerEpoch.leaderAndIsr, controller.epoch))
          // NOTE: the above write can fail only if the current controller lost its zk session and the new controller
          // took over and initialized this partition. This can happen if the current controller went into a long
          // GC pause
          controllerContext.partitionLeadershipInfo.put(topicAndPartition, leaderIsrAndControllerEpoch) // 更新context中的partitionLeadershipInfo
          brokerRequestBatch.addLeaderAndIsrRequestForBrokers(liveAssignedReplicas, topicAndPartition.topic, // 添加LeaderAndIsrRequest到requestbatch
            topicAndPartition.partition, leaderIsrAndControllerEpoch, replicaAssignment)
        } catch {
          case e: ZkNodeExistsException =>
            // read the controller epoch
            val leaderIsrAndEpoch = ReplicationUtils.getLeaderIsrAndEpochForPartition(zkClient, topicAndPartition.topic,
              topicAndPartition.partition).get
            val failMsg = ("encountered error while changing partition %s's state from New to Online since LeaderAndIsr path already " +
                           "exists with value %s and controller epoch %d")
                             .format(topicAndPartition, leaderIsrAndEpoch.leaderAndIsr.toString(), leaderIsrAndEpoch.controllerEpoch)
            stateChangeLogger.error("Controller %d epoch %d ".format(controllerId, controller.epoch) + failMsg)
            throw new StateChangeFailedException(failMsg)
        }
    }
  }

 

OfflinePartition或OnlinePartition –》OnlinePartition

这个相对简单，只需要重新选举一下leader

electLeaderForPartition

def electLeaderForPartition(topic: String, partition: Int, leaderSelector: PartitionLeaderSelector) {
    val topicAndPartition = TopicAndPartition(topic, partition)
    try {
      var zookeeperPathUpdateSucceeded: Boolean = false
      var newLeaderAndIsr: LeaderAndIsr = null
      var replicasForThisPartition: Seq[Int] = Seq.empty[Int]
      while(!zookeeperPathUpdateSucceeded) { // while，只有更新zk成功，或发生异常才会跳出，这样写是不是有点危险
        val currentLeaderIsrAndEpoch = getLeaderIsrAndEpochOrThrowException(topic, partition) // 去zk获取leaderAndIsr信息，如果取不到，抛异常，因为offline或online都应该在zk上有数据的
        val currentLeaderAndIsr = currentLeaderIsrAndEpoch.leaderAndIsr
        val controllerEpoch = currentLeaderIsrAndEpoch.controllerEpoch
        if (controllerEpoch > controller.epoch) { // 判断leaderAndISR如果已经被其他更新epoch的controller改过，那就说明当前controller已经过期了，抛异常
          val failMsg = ("aborted leader election for partition [%s,%d] since the LeaderAndIsr path was " +
                         "already written by another controller. This probably means that the current controller %d went through " +
                         "a soft failure and another controller was elected with epoch %d.")
                           .format(topic, partition, controllerId, controllerEpoch)
          stateChangeLogger.error("Controller %d epoch %d ".format(controllerId, controller.epoch) + failMsg)
          throw new StateChangeFailedException(failMsg)
        }
        // elect new leader or throw exception
        val (leaderAndIsr, replicas) = leaderSelector.selectLeader(topicAndPartition, currentLeaderAndIsr) // 调用Selector来选取leader，不同的Selector会有不同的选取逻辑
        val (updateSucceeded, newVersion) = ReplicationUtils.updateLeaderAndIsr(zkClient, topic, partition, // 在zk上更新leaderAndISR
          leaderAndIsr, controller.epoch, currentLeaderAndIsr.zkVersion)
        newLeaderAndIsr = leaderAndIsr
        newLeaderAndIsr.zkVersion = newVersion
        zookeeperPathUpdateSucceeded = updateSucceeded
        replicasForThisPartition = replicas
      }
      val newLeaderIsrAndControllerEpoch = new LeaderIsrAndControllerEpoch(newLeaderAndIsr, controller.epoch)
      // update the leader cache
      controllerContext.partitionLeadershipInfo.put(TopicAndPartition(topic, partition), newLeaderIsrAndControllerEpoch)
      stateChangeLogger.trace("Controller %d epoch %d elected leader %d for Offline partition %s"
                                .format(controllerId, controller.epoch, newLeaderAndIsr.leader, topicAndPartition))
      val replicas = controllerContext.partitionReplicaAssignment(TopicAndPartition(topic, partition))
      // store new leader and isr info in cache
      brokerRequestBatch.addLeaderAndIsrRequestForBrokers(replicasForThisPartition, topic, partition,
        newLeaderIsrAndControllerEpoch, replicas)
    } catch {
      case lenne: LeaderElectionNotNeededException => // swallow
      case nroe: NoReplicaOnlineException => throw nroe
      case sce: Throwable =>
        val failMsg = "encountered error while electing leader for partition %s due to: %s.".format(topicAndPartition, sce.getMessage)
        stateChangeLogger.error("Controller %d epoch %d ".format(controllerId, controller.epoch) + failMsg)
        throw new StateChangeFailedException(failMsg, sce)
    }
    debug("After leader election, leader cache is updated to %s".format(controllerContext.partitionLeadershipInfo.map(l => (l._1, l._2))))
  }