Apache Kafka源码分析 – Replica and Partition

Replica

对于local replica, 需要记录highWatermarkValue，表示当前已经committed的数据
对于remote replica，需要记录logEndOffsetValue以及更新的时间

package kafka.cluster

class Replica(val brokerId: Int,
              val partition: Partition,
              time: Time = SystemTime,
              initialHighWatermarkValue: Long = 0L,
              val log: Option[Log] = None) extends Logging {
  //only defined in local replica
  private[this] var highWatermarkValue: AtomicLong = new AtomicLong(initialHighWatermarkValue)
  // only used for remote replica; logEndOffsetValue for local replica is kept in log
  private[this] var logEndOffsetValue = new AtomicLong(ReplicaManager.UnknownLogEndOffset)
  private[this] var logEndOffsetUpdateTimeMsValue: AtomicLong = new AtomicLong(time.milliseconds)
  val topic = partition.topic
  val partitionId = partition.partitionId
}

Partition

主要用于管理leader，ISR，AR

package kafka.cluster

/**
 * Data structure that represents a topic partition. The leader maintains the AR, ISR, CUR, RAR
 */
class Partition(val topic: String,
                val partitionId: Int,
                var replicationFactor: Int,
                time: Time,
                val replicaManager: ReplicaManager) extends Logging with KafkaMetricsGroup {
  private val localBrokerId = replicaManager.config.brokerId //当前的brokerId
  private val logManager = replicaManager.logManager
  private val zkClient = replicaManager.zkClient
  var leaderReplicaIdOpt: Option[Int] = None //leaderReplica的id
  var inSyncReplicas: Set[Replica] = Set.empty[Replica] //ISR
  private val assignedReplicaMap = new Pool[Int,Replica] //AR，往往是大于等于ISR的
  private val leaderIsrUpdateLock = new Object
  private var zkVersion: Int = LeaderAndIsr.initialZKVersion
  private var leaderEpoch: Int = LeaderAndIsr.initialLeaderEpoch - 1
  /* Epoch of the controller that last changed the leader. This needs to be initialized correctly upon broker startup.
   * One way of doing that is through the controller's start replica state change command. When a new broker starts up
   * the controller sends it a start replica command containing the leader for each partition that the broker hosts.
   * In addition to the leader, the controller can also send the epoch of the controller that elected the leader for
   * each partition. */
  private var controllerEpoch: Int = KafkaController.InitialControllerEpoch - 1
}

getOrCreateReplica

  def getOrCreateReplica(replicaId: Int = localBrokerId): Replica = {
    val replicaOpt = getReplica(replicaId)
    replicaOpt match {
      case Some(replica) => replica
      case None => //需要create
        if (isReplicaLocal(replicaId)) {
          val config = LogConfig.fromProps(logManager.defaultConfig.toProps, AdminUtils.fetchTopicConfig(zkClient, topic))
          val log = logManager.createLog(TopicAndPartition(topic, partitionId), config) //创建logfile
          val checkpoint = replicaManager.highWatermarkCheckpoints(log.dir.getParent) //试图读取HW的checkpoint
          val offsetMap = checkpoint.read
          if (!offsetMap.contains(TopicAndPartition(topic, partitionId)))
            warn("No checkpointed highwatermark is found for partition [%s,%d]".format(topic, partitionId))
          val offset = offsetMap.getOrElse(TopicAndPartition(topic, partitionId), 0L).min(log.logEndOffset) //offset为cp读出的HW和logEndOffset中小的那个
          val localReplica = new Replica(replicaId, this, time, offset, Some(log)) //创建Replica对象
          addReplicaIfNotExists(localReplica) //加到AR中去
        } else { //Remote Replica，直接创建对象
          val remoteReplica = new Replica(replicaId, this, time)
          addReplicaIfNotExists(remoteReplica)
        }
        getReplica(replicaId).get
    }
  }

makeLeader

  /**
   * Make the local replica the leader by resetting LogEndOffset for remote replicas (there could be old LogEndOffset from the time when this broker was the leader last time)
   *  and setting the new leader and ISR
   */
  def makeLeader(controllerId: Int,
                 partitionStateInfo: PartitionStateInfo, correlationId: Int): Boolean = {
    leaderIsrUpdateLock synchronized {
      val allReplicas = partitionStateInfo.allReplicas  //取出所有replicaid
      val leaderIsrAndControllerEpoch = partitionStateInfo.leaderIsrAndControllerEpoch
      val leaderAndIsr = leaderIsrAndControllerEpoch.leaderAndIsr
      // record the epoch of the controller that made the leadership decision. This is useful while updating the isr
      // to maintain the decision maker controller's epoch in the zookeeper path
      controllerEpoch = leaderIsrAndControllerEpoch.controllerEpoch //每次做leadership decision的时候需要更新controllerEpoch
      // add replicas that are new
      allReplicas.foreach(replica => getOrCreateReplica(replica)) //生成所有replica对象，对于new的需要create
      val newInSyncReplicas = leaderAndIsr.isr.map(r => getOrCreateReplica(r)).toSet //生成新的ISR
      // remove assigned replicas that have been removed by the controller
      (assignedReplicas().map(_.brokerId) -- allReplicas).foreach(removeReplica(_)) //对于AR中已经不再有效的replica，remove
      // reset LogEndOffset for remote replicas
      assignedReplicas.foreach(r => if (r.brokerId != localBrokerId) r.logEndOffset = ReplicaManager.UnknownLogEndOffset) //清空所有remote replica的LEO
      inSyncReplicas = newInSyncReplicas
      leaderEpoch = leaderAndIsr.leaderEpoch
      zkVersion = leaderAndIsr.zkVersion
      leaderReplicaIdOpt = Some(localBrokerId)
      // we may need to increment high watermark
      maybeIncrementLeaderHW(getReplica().get) //更新HW，用min（所有remote replica的LEO）
      true
    }
  }

maybeIncrementLeaderHW
用所有remote replica的LEO的最小值来替换当前的HW（如果大于HW的话）

  /**
   * There is no need to acquire the leaderIsrUpdate lock here since all callers of this private API acquire that lock
   * @param leaderReplica
   */
  private def maybeIncrementLeaderHW(leaderReplica: Replica) {
    val allLogEndOffsets = inSyncReplicas.map(_.logEndOffset)
    val newHighWatermark = allLogEndOffsets.min
    val oldHighWatermark = leaderReplica.highWatermark
    if(newHighWatermark > oldHighWatermark) {
      leaderReplica.highWatermark = newHighWatermark //更新HW
    }
    else
  }

makeFollower

  /**
   *  Make the local replica the follower by setting the new leader and ISR to empty
   */
  def makeFollower(controllerId: Int,
                   partitionStateInfo: PartitionStateInfo,
                   leaders: Set[Broker], correlationId: Int): Boolean = {
    leaderIsrUpdateLock synchronized {
      val allReplicas = partitionStateInfo.allReplicas
      val leaderIsrAndControllerEpoch = partitionStateInfo.leaderIsrAndControllerEpoch
      val leaderAndIsr = leaderIsrAndControllerEpoch.leaderAndIsr
      val newLeaderBrokerId: Int = leaderAndIsr.leader //新的leaderid
      // record the epoch of the controller that made the leadership decision. This is useful while updating the isr
      // to maintain the decision maker controller's epoch in the zookeeper path
      controllerEpoch = leaderIsrAndControllerEpoch.controllerEpoch
      // TODO: Delete leaders from LeaderAndIsrRequest in 0.8.1
      leaders.find(_.id == newLeaderBrokerId) match {
        case Some(leaderBroker) =>
          // add replicas that are new
          allReplicas.foreach(r => getOrCreateReplica(r))
          // remove assigned replicas that have been removed by the controller
          (assignedReplicas().map(_.brokerId) -- allReplicas).foreach(removeReplica(_))
          inSyncReplicas = Set.empty[Replica] //将ISR置空
          leaderEpoch = leaderAndIsr.leaderEpoch
          zkVersion = leaderAndIsr.zkVersion
          leaderReplicaIdOpt = Some(newLeaderBrokerId)
        case None => // we should not come here
      }
      true
    }
  }

maybeShrinkIsr

从ISR中将Stuck followers和Slow followers去除

  def maybeShrinkIsr(replicaMaxLagTimeMs: Long,  replicaMaxLagMessages: Long) {
    leaderIsrUpdateLock synchronized {
      leaderReplicaIfLocal() match {
        case Some(leaderReplica) =>
          val outOfSyncReplicas = getOutOfSyncReplicas(leaderReplica, replicaMaxLagTimeMs, replicaMaxLagMessages) // 获取OutofSync的replicas
          if(outOfSyncReplicas.size > 0) {
            val newInSyncReplicas = inSyncReplicas – outOfSyncReplicas //从ISR中去除outOfSyncReplicas
            // update ISR in zk and in cache
            updateIsr(newInSyncReplicas)
            // we may need to increment high watermark since ISR could be down to 1
            maybeIncrementLeaderHW(leaderReplica)
            replicaManager.isrShrinkRate.mark()
          }
        case None => // do nothing if no longer leader
      }
    }
  }

  def getOutOfSyncReplicas(leaderReplica: Replica, keepInSyncTimeMs: Long, keepInSyncMessages: Long): Set[Replica] = {
    /**
     * there are two cases that need to be handled here -
     * 1. Stuck followers: If the leo of the replica hasn't been updated for keepInSyncTimeMs ms,
     *                     the follower is stuck and should be removed from the ISR
     * 2. Slow followers: If the leo of the slowest follower is behind the leo of the leader by keepInSyncMessages, the
     *                     follower is not catching up and should be removed from the ISR
     **/
    val leaderLogEndOffset = leaderReplica.logEndOffset
    val candidateReplicas = inSyncReplicas - leaderReplica
    // Case 1 above
    val stuckReplicas = candidateReplicas.filter(r => (time.milliseconds - r.logEndOffsetUpdateTimeMs) > keepInSyncTimeMs)
    if(stuckReplicas.size > 0)
      debug("Stuck replicas for partition [%s,%d] are %s".format(topic, partitionId, stuckReplicas.map(_.brokerId).mkString(",")))
    // Case 2 above
    val slowReplicas = candidateReplicas.filter(r => r.logEndOffset >= 0 && (leaderLogEndOffset - r.logEndOffset) > keepInSyncMessages)
    if(slowReplicas.size > 0)
      debug("Slow replicas for partition [%s,%d] are %s".format(topic, partitionId, slowReplicas.map(_.brokerId).mkString(",")))
    stuckReplicas ++ slowReplicas
  }