Kafka原始碼分析

rilley發表於2016-04-14
Kafka原始碼

 

本文主要針對於Kafka的原始碼進行分析，版本為kafka-0.8.2.1。 由於時間有限，可能更新比較慢...

 

 

Kafka.scala

 

// 讀取配置檔案
val props = Utils.loadProps(args(0))
val serverConfig = new KafkaConfig(props)
KafkaMetricsReporter.startReporters(serverConfig.props)

val kafkaServerStartable = new KafkaServerStartable(serverConfig)

// 註冊一個關閉鉤子，當JVM關閉時呼叫KafkaServerStartable.shutdown 
Runtime.getRuntime().addShutdownHook(new Thread() {
  override def run() = kafkaServerStartable.shutdown
})  

// 執行並等待結束
kafkaServerStartable.startup
kafkaServerStartable.awaitShutdown

　　

 

Server

 

        實際呼叫類為KafkaServer

def startup() {
  kafkaScheduler.startup()

  // 初始化Zookeeper內相關路徑
  zkClient = initZk()

  // 日誌管理器
  logManager = createLogManager(zkClient, brokerState)
  logManager.startup()

  socketServer = new SocketServer(...)
  socketServer.startup()

  // 啟動副本管理器
  replicaManager = new ReplicaManager(config, time, zkClient, kafkaScheduler, logManager, isShuttingDown)

  // 建立偏移量管理器
  offsetManager = createOffsetManager()

  // 例項化排程器
  kafkaController = new KafkaController(config, zkClient, brokerState)

  // 請求處理器
  apis = new KafkaApis(...)

  // 網路請求處理
  requestHandlerPool = new KafkaRequestHandlerPool(config.brokerId, socketServer.requestChannel, apis, config.numIoThreads)
  brokerState.newState(RunningAsBroker)

  Mx4jLoader.maybeLoad()
  replicaManager.startup()
  kafkaController.startup()

  // Topic配置管理器
  topicConfigManager = new TopicConfigManager(zkClient, logManager)
  topicConfigManager.startup()

  // Broker的心跳檢查
  kafkaHealthcheck = new KafkaHealthcheck(...)
  kafkaHealthcheck.startup()

  registerStats()
  startupComplete.set(true)
  info("started")
}

 

        在KafkaServer的startup中看到主要進行幾個主要服務的初始化和啟動。

 

private def initZk(): ZkClient =
{
  info("Connecting to zookeeper on " + config.zkConnect)
   
  // Kafka在Zookeeper中的工作根目錄
  val chroot = {
    if (config.zkConnect.indexOf("/") > 0)
      config.zkConnect.substring(config.zkConnect.indexOf("/"))
    else
      ""
  }
  // 建立工作根目錄
  if (chroot.length > 1) {
    val zkConnForChrootCreation = config.zkConnect.substring(0, config.zkConnect.indexOf("/"))
    val zkClientForChrootCreation = new ZkClient(...)
    ZkUtils.makeSurePersistentPathExists(zkClientForChrootCreation, chroot)
    info("Created zookeeper path " + chroot)
    zkClientForChrootCreation.close()
  }
 
  // 例項化ZkClient
  val zkClient = new ZkClient(config.zkConnect, config.zkSessionTimeoutMs, config.zkConnectionTimeoutMs, ZKStringSerializer)
  // 在Zookeeper中建立必要持久路徑
  ZkUtils.setupCommonPaths(zkClient)
  zkClient
}

　　

        KafkaScheduler實際為對執行緒池ScheduledThreadPoolExecutor的封裝，這裡不做過多的分析。

 

KafkaHealthcheck(...)
{
  val brokerIdPath = ZkUtils.BrokerIdsPath + "/" + brokerId
  val sessionExpireListener = new SessionExpireListener

  def startup() 
  {
    // 註冊一個Zookeeper事件(狀態)監聽器
    zkClient.subscribeStateChanges(sessionExpireListener)
    // 在Zookeeper的/brokers/ids/id目錄建立臨時節點並寫入節點資訊   
    register()
  }
}

       

        IZkStateListener 定義了兩種事件：一種是連線狀態的改變，例如由未連線改變成連線上，連線上改為過期等；

        另一種建立一個新的session（連線）， 通常是由於session失效然後新的session被建立時觸發。

class SessionExpireListener() extends IZkStateListener 
{
  @throws(classOf[Exception])
  def handleStateChanged(state: KeeperState) {}

  @throws(classOf[Exception])
  def handleNewSession() = register()
}

　　

 ReplicaManager

def startup() 
{
  scheduler.schedule("isr-expiration", maybeShrinkIsr, period = config.replicaLagTimeMaxMs, unit = TimeUnit.MILLISECONDS)
}
// 定時呼叫maybeShrinkIsr
private def maybeShrinkIsr(): Unit = 
{
  trace("Evaluating ISR list of partitions to see which replicas can be removed from the ISR")
  allPartitions.values.foreach(partition => partition.maybeShrinkIsr(config.replicaLagTimeMaxMs, config.replicaLagMaxMessages))
}

　　

這裡呼叫了cluster.Partition中的maybeShrinkIsr來將卡住的或者低效的副本從ISR中去除並更新HighWatermark。

 

def maybeShrinkIsr(replicaMaxLagTimeMs: Long,  replicaMaxLagMessages: Long) 
{                        
  inWriteLock(leaderIsrUpdateLock) {                                                                 
    leaderReplicaIfLocal() match {                                                                   
      case Some(leaderReplica) =>                                                                    
      // 找出卡住和低效的Replica並從ISR中去除
        val outOfSyncReplicas = getOutOfSyncReplicas(leaderReplica, replicaMaxLagTimeMs, replicaMaxLagMessages)   
        if(outOfSyncReplicas.size > 0) {
          val newInSyncReplicas = inSyncReplicas -- outOfSyncReplicas                                
          assert(newInSyncReplicas.size > 0)
          // 更新ZK中的ISR                                                       
          updateIsr(newInSyncReplicas)  
          // 計算HW並更新
          maybeIncrementLeaderHW(leaderReplica)
          replicaManager.isrShrinkRate.mark()                                                        
        }                                          
  ...
}

 

def getOutOfSyncReplicas(leaderReplica: Replica, keepInSyncTimeMs: Long, keepInSyncMessages: Long): Set[Replica] = 
{
  // Leader的最後寫入偏移量
  val leaderLogEndOffset = leaderReplica.logEndOffset
  // ISR中排除LeaderReplica的其他集合
  val candidateReplicas = inSyncReplicas - leaderReplica
  // 卡住的Replica集合
  val stuckReplicas = candidateReplicas.filter(r => (time.milliseconds - r.logEndOffsetUpdateTimeMs) > keepInSyncTimeMs)
  // 低效的Replica
  // 條件1 Replicas的offset > 0
  // 條件2 Leader的offset - Replicas的offset > 閥值
  val slowReplicas = candidateReplicas.filter(r =>
    r.logEndOffset.messageOffset >= 0 &&
    leaderLogEndOffset.messageOffset - r.logEndOffset.messageOffset > keepInSyncMessages)
  // 返回卡住的和低效的Replicas
  stuckReplicas ++ slowReplicas
}

　　

Cluster

 

 

Controller

 

 

相關文章