Kafka 副本管理模块（三）：ReplicaManager 类介绍

最新推荐文章于 2022-01-22 23:37:39 发布

其实系一个须刨

最新推荐文章于 2022-01-22 23:37:39 发布

阅读量358

点赞数

分类专栏： kafka-2.4.1

本文链接：https://blog.csdn.net/lianggx3/article/details/108902004

版权

kafka-2.4.1 专栏收录该内容

55 篇文章 19 订阅

订阅专栏

ReplicaManager 负责管理和操作集群中 Broker 的副本，它定义如下：

class ReplicaManager(val config: KafkaConfig, // 配置管理类
                     metrics: Metrics, // 监控指标类
                     time: Time,  // 定时器类
                     val zkClient: KafkaZkClient, // ZooKeeper客户端
                     scheduler: Scheduler,   // Kafka调度器
                     val logManager: LogManager,
                     val isShuttingDown: AtomicBoolean, // 是否已经关闭
                     quotaManagers: QuotaManagers,  // 配额管理器
                     val brokerTopicStats: BrokerTopicStats, // Broker主题监控指标类
                     val metadataCache: MetadataCache,  // Broker元数据缓存，保存集群上分区的 Leader、ISR 等信息。
                     logDirFailureChannel: LogDirFailureChannel,
                     val delayedProducePurgatory: DelayedOperationPurgatory[DelayedProduce], // 处理延时PRODUCE请求的Purgatory
                     val delayedFetchPurgatory: DelayedOperationPurgatory[DelayedFetch],  // 处理延时FETCH请求的Purgatory
                     val delayedDeleteRecordsPurgatory: DelayedOperationPurgatory[DelayedDeleteRecords],  // 处理延时DELETE_RECORDS请求的Purgatory
                     val delayedElectLeaderPurgatory: DelayedOperationPurgatory[DelayedElectLeader], // 处理延时ELECT_LEADERS请求的Purgatory
                     threadNamePrefix: Option[String]) extends Logging with KafkaMetricsGroup {
  /* epoch of the controller that last changed the leader */
  // 作用是隔离过期 Controller 发送的请求。老的 Controller 发送的请求不能再被继续处理了
  // 该字段表示最新一次变更分区 Leader 的 Controller 的 Epoch 值，其默认值为 0。Controller 每发生一次变更，该字段值都会 +1。
  @volatile var controllerEpoch: Int = KafkaController.InitialControllerEpoch
  /** 本地 broker 的 ID */
  private val localBrokerId = config.brokerId
  // Broker 上保存的所有分区对象数据
  private val allPartitions = new Pool[TopicPartition, HostedPartition](
    valueFactory = Some(tp => HostedPartition.Online(Partition(tp, time, this)))
  )
  private val replicaStateChangeLock = new Object
  // replicaFetcherManager 它的主要任务是创建 ReplicaFetcherThread 类实例。
  val replicaFetcherManager = createReplicaFetcherManager(metrics, time, threadNamePrefix, quotaManagers.follower)
  val replicaAlterLogDirsManager = createReplicaAlterLogDirsManager(quotaManagers.alterLogDirs, brokerTopicStats)
  /** 标记 highwatermark-checkpoint 定时任务是否已经启动 */
  private val highWatermarkCheckPointThreadStarted = new AtomicBoolean(false)
  /** 记录每个 log 目录与对应 topic 分区 HW 值的映射关系 */
  @volatile var highWatermarkCheckpoints: Map[String, OffsetCheckpointFile] = logManager.liveLogDirs.map(dir =>
    (dir.getAbsolutePath, new OffsetCheckpointFile(new File(dir, ReplicaManager.HighWatermarkFilename), logDirFailureChannel))).toMap

  this.logIdent = s"[ReplicaManager broker=$localBrokerId] "
  private val stateChangeLogger = new StateChangeLogger(localBrokerId, inControllerContext = false, None)
  /** 记录 ISR 集合发生变化的 topic 分区信息 */
  private val isrChangeSet: mutable.Set[TopicPartition] = new mutable.HashSet[TopicPartition]()
  private val lastIsrChangeMs = new AtomicLong(System.currentTimeMillis())
  private val lastIsrPropagationMs = new AtomicLong(System.currentTimeMillis())

  private var logDirFailureHandler: LogDirFailureHandler = null
}

其中一个重要的字段是replicaFetcherManager，它的主要任务是创建 ReplicaFetcherThread 类实例。上介绍了 ReplicaFetcherThread 类，它的主要职责是帮助 Follower 副本向 Leader 副本拉取消息，并写入到本地日志中。它的定义和主要方法如下:

class ReplicaFetcherManager(brokerConfig: KafkaConfig,
                            protected val replicaManager: ReplicaManager,
                            metrics: Metrics,
                            time: Time,
                            threadNamePrefix: Option[String] = None,
                            quotaManager: ReplicationQuotaManager)
      extends AbstractFetcherManager[ReplicaFetcherThread](
        name = "ReplicaFetcherManager on broker " + brokerConfig.brokerId,
        clientId = "Replica",
        numFetchers = brokerConfig.numReplicaFetchers) {

  override def createFetcherThread(fetcherId: Int, sourceBroker: BrokerEndPoint): ReplicaFetcherThread = {
    val prefix = threadNamePrefix.map(tp => s"$tp:").getOrElse("")
    val threadName = s"${prefix}ReplicaFetcherThread-$fetcherId-${sourceBroker.id}"
    // 创建ReplicaFetcherThread线程实例并返回
    new ReplicaFetcherThread(threadName, fetcherId, sourceBroker, brokerConfig, failedPartitions, replicaManager,
      metrics, time, quotaManager)
  }
}

Kafka 服务在启动时会创建 ReplicaManager 对象，并调用 ReplicaManager#startup 方法启动 ReplicaManager 管理的定时任务，即 isr-expiration 和 isr-change-propagation 定时任务。

  def startup(): Unit = {
    // start ISR expiration thread
    // A follower can lag behind leader for up to config.replicaLagTimeMaxMs x 1.5 before it is removed from ISR
    // 定时检测当前 broker 节点管理的每个分区是否需要缩减 ISR 集合，并执行缩减操作
    // 周期性执行 ReplicaManager#maybeShrinkIsr 方法，尝试缩减当前 broker 节点管理的分区对应的 ISR 集合，具体缩减操作由 Partition#maybeShrinkIsr 方法实现
    scheduler.schedule("isr-expiration", maybeShrinkIsr _, period = config.replicaLagTimeMaxMs / 2, unit = TimeUnit.MILLISECONDS)
    // 定时将 ISR 集合发生变化的 topic 分区记录到 ZK，这个会在指挥详细介绍
    scheduler.schedule("isr-change-propagation", maybePropagateIsrChanges _, period = 2500L, unit = TimeUnit.MILLISECONDS)
    scheduler.schedule("shutdown-idle-replica-alter-log-dirs-thread", shutdownIdleReplicaAlterLogDirsThread _, period = 10000L, unit = TimeUnit.MILLISECONDS)

    // If inter-broker protocol (IBP) < 1.0, the controller will send LeaderAndIsrRequest V0 which does not include isNew field.
    // In this case, the broker receiving the request cannot determine whether it is safe to create a partition if a log directory has failed.
    // Thus, we choose to halt the broker on any log diretory failure if IBP < 1.0
    val haltBrokerOnFailure = config.interBrokerProtocolVersion < KAFKA_1_0_IV0
    logDirFailureHandler = new LogDirFailureHandler("LogDirFailureHandler", haltBrokerOnFailure)
    logDirFailureHandler.start()
  }