Spark 运行架构核心总结

摘要：

1.基本术语

2.运行架构

2.1基本架构

2.2运行流程

　  2.3相关的UML类图

　  2.4调度模块：

2.4.1作业调度简介

2.4.2任务调度简介

3.运行模式

3.1 standalone模式

4.RDD实战

总结：

1. 基本术语：

• Application：在Spark 上建立的用户程序，一个程序由一个驱动程序（Driver Program）和集群中的执行进程（Executer）构成。
• Driver Program：运行应用程序（Application）的main函数和创建SparkContext的程序。
• Executer：运行在工作节点（Work Node）上的进程。Executer负责运行任务（Task）并将各节点的数据保存在内存或磁盘中。每个应用程序都有自己对应Executer
• Work Node:集群中运行应用程序（Applicatuon）的节点
• Cluster Manager: 在集群上获取资源的外部服务（如Standalone,Mesos,Yarn)，称作资源管理器或集群管理器
• Job: 包含多个Task的并行计算，往往由Spark Action（如save,collect）触发生成，一个Application中往往会产生多个Job
• Stage:每个Job被分成了更小的任务集合（TaskSet），各个阶段（Stage）相互依赖
• Task:被发送到某一个Executer的工作单元
• DAGScheduler:基于Stage的逻辑调度模块,负责将每个Job分割成一个DAG图
• TaskScheduler:基于Task的任务调度模块，负责每个Task的跟踪和向DAGScheduler汇报任务执行情况

　　2.运行架构

　　　　2.1基本架构：

　　图示：

　　　　　

　　　　Spark Application在集群中以一组独立的进程运行，通过你的驱动程序（driver program）中的SparkContext 对象进行协作。

　　　　具体来说，SparkContext可以连接到多种类型的集群管理器 cluster managers  （standalone cluster manager,  Mesos ,YARN），这些 cluster managers 负责跨应用程序分配资源。一旦连接，Spark就获得集群中的节点上的executors，接下来，它会将应用程序代码发送到executors。最后，SparkContext发送tasks到executors运行。

　　　　注意：该驱动程序会一直监听并接受其executor传入的连接（spark.driver.port在网络配置部分）。这样，driver program必须可以寻找到工作节点的网络地址。数据不能跨应用程序（SparkContext）访问，除非写入外部系统

2.1.1 SparkContext类（代表连接到spark集群，现在一个jvm只能有一个sc,以后会取消）：

　　　　几个重要的属性（包含DAGScheduler,TaskScheduler调度，获取executor,心跳与监听等）：

　　　　说明：这里的下划线_代表默认值，比如Int 默认值就是0,String默认值就是None  参考知乎

  /* ------------------------------------------------------------------------------------- *
   | Private variables. These variables keep the internal state of the context, and are    |
   | not accessible by the outside world. They're mutable since we want to initialize all  |
   | of them to some neutral value ahead of time, so that calling "stop()" while the       |
   | constructor is still running is safe.                                                 |
   * ------------------------------------------------------------------------------------- */

  private var _conf: SparkConf = _
  private var _eventLogDir: Option[URI] = None
  private var _eventLogCodec: Option[String] = None
  private var _env: SparkEnv = _
  private var _jobProgressListener: JobProgressListener = _
  private var _statusTracker: SparkStatusTracker = _
  private var _progressBar: Option[ConsoleProgressBar] = None
  private var _ui: Option[SparkUI] = None
  private var _hadoopConfiguration: Configuration = _
  private var _executorMemory: Int = _
  private var _schedulerBackend: SchedulerBackend = _
  private var _taskScheduler: TaskScheduler = _
  private var _heartbeatReceiver: RpcEndpointRef = _
  @volatile private var _dagScheduler: DAGScheduler = _
  private var _applicationId: String = _
  private var _applicationAttemptId: Option[String] = None
  private var _eventLogger: Option[EventLoggingListener] = None
  private var _executorAllocationManager: Option[ExecutorAllocationManager] = None
  private var _cleaner: Option[ContextCleaner] = None
  private var _listenerBusStarted: Boolean = false
  private var _jars: Seq[String] = _
  private var _files: Seq[String] = _
  private var _shutdownHookRef: AnyRef = _

2.1.2 Executor（一个运行任务的线程池，通过RPC与Driver通信）：

　心跳报告（心跳进程，记录心跳失败次数和接受task的心跳）：

　这里有两个参数：spark.executor.heartbeat.maxFailures = 60，spark.executor.heartbeatInterval = 10s，意味着最多每隔10min会重新发送一次心跳

 

  /** Reports heartbeat and metrics for active tasks to the driver. */
  private def reportHeartBeat(): Unit = {
    // list of (task id, accumUpdates) to send back to the driver
    val accumUpdates = new ArrayBuffer[(Long, Seq[AccumulatorV2[_, _]])]()
    val curGCTime = computeTotalGcTime()

    for (taskRunner <- runningTasks.values().asScala) {
      if (taskRunner.task != null) {
        taskRunner.task.metrics.mergeShuffleReadMetrics()
        taskRunner.task.metrics.setJvmGCTime(curGCTime - taskRunner.startGCTime)
        accumUpdates += ((taskRunner.taskId, taskRunner.task.metrics.accumulators()))
      }
    }

    val message = Heartbeat(executorId, accumUpdates.toArray, env.blockManager.blockManagerId)
    try {
      val response = heartbeatReceiverRef.askWithRetry[HeartbeatResponse](
          message, RpcTimeout(conf, "spark.executor.heartbeatInterval", "10s"))
      if (response.reregisterBlockManager) {
        logInfo("Told to re-register on heartbeat")
        env.blockManager.reregister()
      }
      heartbeatFailures = 0
    } catch {
      case NonFatal(e) =>
        logWarning("Issue communicating with driver in heartbeater", e)
        heartbeatFailures += 1
        if (heartbeatFailures >= HEARTBEAT_MAX_FAILURES) {
          logError(s"Exit as unable to send heartbeats to driver " +
            s"more than $HEARTBEAT_MAX_FAILURES times")
          System.exit(ExecutorExitCode.HEARTBEAT_FAILURE)
        }
    }
  }

　　

Task管理（taskRunner类的启动，停止）

  // Maintains the list of running tasks.
  private val runningTasks = new ConcurrentHashMap[Long, TaskRunner]

下面是TaskRunner 的run方法，贴出来，以后研究

    override def run(): Unit = {
      val taskMemoryManager = new TaskMemoryManager(env.memoryManager, taskId)
      val deserializeStartTime = System.currentTimeMillis()
      Thread.currentThread.setContextClassLoader(replClassLoader)
      val ser = env.closureSerializer.newInstance()
      logInfo(s"Running $taskName (TID $taskId)")
      execBackend.statusUpdate(taskId, TaskState.RUNNING, EMPTY_BYTE_BUFFER)
      var taskStart: Long = 0
      startGCTime = computeTotalGcTime()

      try {
        val (taskFiles, taskJars, taskProps, taskBytes) =
          Task.deserializeWithDependencies(serializedTask)

        // Must be set before updateDependencies() is called, in case fetching dependencies
        // requires access to properties contained within (e.g. for access control).
        Executor.taskDeserializationProps.set(taskProps)

        updateDependencies(taskFiles, taskJars)
        task = ser.deserialize[Task[Any]](taskBytes, Thread.currentThread.getContextClassLoader)
        task.localProperties = taskProps
        task.setTaskMemoryManager(taskMemoryManager)

        // If this task has been killed before we deserialized it, let's quit now. Otherwise,
        // continue executing the task.
        if (killed) {
          // Throw an exception rather than returning, because returning within a try{} block
          // causes a NonLocalReturnControl exception to be thrown. The NonLocalReturnControl
          // exception will be caught by the catch block, leading to an incorrect ExceptionFailure
          // for the task.
          throw new TaskKilledException
        }

        logDebug("Task " + taskId + "'s epoch is " + task.epoch)
        env.mapOutputTracker.updateEpoch(task.epoch)

        // Run the actual task and measure its runtime.
        taskStart = System.currentTimeMillis()
        var threwException = true
        val value = try {
          val res = task.run(
            taskAttemptId = taskId,
            attemptNumber = attemptNumber,
            metricsSystem = env.metricsSystem)
          threwException = false
          res
        } finally {
          val releasedLocks = env.blockManager.releaseAllLocksForTask(taskId)
          val freedMemory = taskMemoryManager.cleanUpAllAllocatedMemory()

          if (freedMemory > 0) {
            val errMsg = s"Managed memory leak detected; size = $freedMemory bytes, TID = $taskId"
            if (conf.getBoolean("spark.unsafe.exceptionOnMemoryLeak", false) && !threwException) {
              throw new SparkException(errMsg)
            } else {
              logError(errMsg)
            }
          }

          if (releasedLocks.nonEmpty) {
            val errMsg =
              s"${releasedLocks.size} block locks were not released by TID = $taskId:\n" +
                releasedLocks.mkString("[", ", ", "]")
            if (conf.getBoolean("spark.storage.exceptionOnPinLeak", false) && !threwException) {
              throw new SparkException(errMsg)
            } else {
              logWarning(errMsg)
            }
          }
        }
        val taskFinish = System.currentTimeMillis()

        // If the task has been killed, let's fail it.
        if (task.killed) {
          throw new TaskKilledException
        }

        val resultSer = env.serializer.newInstance()
        val beforeSerialization = System.currentTimeMillis()
        val valueBytes = resultSer.serialize(value)
        val afterSerialization = System.currentTimeMillis()

        // Deserialization happens in two parts: first, we deserialize a Task object, which
        // includes the Partition. Second, Task.run() deserializes the RDD and function to be run.
        task.metrics.setExecutorDeserializeTime(
          (taskStart - deserializeStartTime) + task.executorDeserializeTime)
        // We need to subtract Task.run()'s deserialization time to avoid double-counting
        task.metrics.setExecutorRunTime((taskFinish - taskStart) - task.executorDeserializeTime)
        task.metrics.setJvmGCTime(computeTotalGcTime() - startGCTime)
        task.metrics.setResultSerializationTime(afterSerialization - beforeSerialization)

        // Note: accumulator updates must be collected after TaskMetrics is updated
        val accumUpdates = task.collectAccumulatorUpdates()
        // TODO: do not serialize value twice
        val directResult = new DirectTaskResult(valueBytes, accumUpdates)
        val serializedDirectResult = ser.serialize(directResult)
        val resultSize = serializedDirectResult.limit

        // directSend = sending directly back to the driver
        val serializedResult: ByteBuffer = {
          if (maxResultSize > 0 && resultSize > maxResultSize) {
            logWarning(s"Finished $taskName (TID $taskId). Result is larger than maxResultSize " +
              s"(${Utils.bytesToString(resultSize)} > ${Utils.bytesToString(maxResultSize)}), " +
              s"dropping it.")
            ser.serialize(new IndirectTaskResult[Any](TaskResultBlockId(taskId), resultSize))
          } else if (resultSize > maxDirectResultSize) {
            val blockId = TaskResultBlockId(taskId)
            env.blockManager.putBytes(
              blockId,
              new ChunkedByteBuffer(serializedDirectResult.duplicate()),
              StorageLevel.MEMORY_AND_DISK_SER)
            logInfo(
              s"Finished $taskName (TID $taskId). $resultSize bytes result sent via BlockManager)")
            ser.serialize(new IndirectTaskResult[Any](blockId, resultSize))
          } else {
            logInfo(s"Finished $taskName (TID $taskId). $resultSize bytes result sent to driver")
            serializedDirectResult
          }
        }

        execBackend.statusUpdate(taskId, TaskState.FINISHED, serializedResult)

      } catch {
        case ffe: FetchFailedException =>
          val reason = ffe.toTaskEndReason
          setTaskFinishedAndClearInterruptStatus()
          execBackend.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))

        case _: TaskKilledException | _: InterruptedException if task.killed =>
          logInfo(s"Executor killed $taskName (TID $taskId)")
          setTaskFinishedAndClearInterruptStatus()
          execBackend.statusUpdate(taskId, TaskState.KILLED, ser.serialize(TaskKilled))

        case CausedBy(cDE: CommitDeniedException) =>
          val reason = cDE.toTaskEndReason
          setTaskFinishedAndClearInterruptStatus()
          execBackend.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))

        case t: Throwable =>
          // Attempt to exit cleanly by informing the driver of our failure.
          // If anything goes wrong (or this was a fatal exception), we will delegate to
          // the default uncaught exception handler, which will terminate the Executor.
          logError(s"Exception in $taskName (TID $taskId)", t)

          // Collect latest accumulator values to report back to the driver
          val accums: Seq[AccumulatorV2[_, _]] =
            if (task != null) {
              task.metrics.setExecutorRunTime(System.currentTimeMillis() - taskStart)
              task.metrics.setJvmGCTime(computeTotalGcTime() - startGCTime)
              task.collectAccumulatorUpdates(taskFailed = true)
            } else {
              Seq.empty
            }

          val accUpdates = accums.map(acc => acc.toInfo(Some(acc.value), None))

          val serializedTaskEndReason = {
            try {
              ser.serialize(new ExceptionFailure(t, accUpdates).withAccums(accums))
            } catch {
              case _: NotSerializableException =>
                // t is not serializable so just send the stacktrace
                ser.serialize(new ExceptionFailure(t, accUpdates, false).withAccums(accums))
            }
          }
          setTaskFinishedAndClearInterruptStatus()
          execBackend.statusUpdate(taskId, TaskState.FAILED, serializedTaskEndReason)

          // Don't forcibly exit unless the exception was inherently fatal, to avoid
          // stopping other tasks unnecessarily.
          if (Utils.isFatalError(t)) {
            SparkUncaughtExceptionHandler.uncaughtException(t)
          }

      } finally {
        runningTasks.remove(taskId)
      }
    }

　　　

　　

　　　　2.2运行流程:

　　　　　　图示：

　　　　　　

　　　　　　注意这里的StandaloneExecutorBackend是一个概念（我在spark项目中没找到），实际上的spark standalone的资源调度类是 CoarseGrainedExecutorBackend

　　　　　　1.构建Spark Application的运行环境（启动SparkContext），SparkContext向资源管理器（ClusterManager）（可以是Standalone、Mesos或YARN）注册并申请运行Executor资源；
　　　　　　2.资源管理器分配Executor资源并启动StandaloneExecutorBackend，Executor运行情况将随着心跳发送到资源管理器上； 
　　　　　　3.SparkContext构建成DAG图，将DAG图分解成Stage，并把Taskset发送给Task Scheduler。

　　　　　　　Executor向SparkContext申请Task，Task Scheduler将Task发放给Executor运行同时SparkContext将应用程序代码发放给Executor。

　　　　　　4.Task在Executor上运行，运行完毕释放所有资源。

　　　　　2.3相关的类：　　　

　　　　　　　　ExecutorBackend:

　　　　　　　　特质签名（Executor用来向集群调度发送更新的插件）

　　　　　　　　

　　　　　　　　各种运行模式的类图：

　　　　　　　　

　　　　　　其中standalone是用SparkDeploySchedulerBackend配合TeskSchedulerImpl工作，相关类图应该是：

　　　　　　

　　　　　　SchedulerBackend特质（核心函数：reviveOffers()）

　　　　　　　

CoarseGrainedExecutorBackend（receive方法里是若干模式匹配，类似于switch case，根据相关模式执行相应操作。主要有注册Executor,运行Task等）

override def receive: PartialFunction[Any, Unit] = {
    case RegisteredExecutor(hostname) =>
      logInfo("Successfully registered with driver")
      executor = new Executor(executorId, hostname, env, userClassPath, isLocal = false)

    case RegisterExecutorFailed(message) =>
      logError("Slave registration failed: " + message)
      exitExecutor(1)

    case LaunchTask(data) =>
      if (executor == null) {
        logError("Received LaunchTask command but executor was null")
        exitExecutor(1)
      } else {
        val taskDesc = ser.deserialize[TaskDescription](data.value)
        logInfo("Got assigned task " + taskDesc.taskId)
        executor.launchTask(this, taskId = taskDesc.taskId, attemptNumber = taskDesc.attemptNumber,
          taskDesc.name, taskDesc.serializedTask)
      }

    case KillTask(taskId, _, interruptThread) =>
      if (executor == null) {
        logError("Received KillTask command but executor was null")
        exitExecutor(1)
      } else {
        executor.killTask(taskId, interruptThread)
      }

    case StopExecutor =>
      stopping.set(true)
      logInfo("Driver commanded a shutdown")
      // Cannot shutdown here because an ack may need to be sent back to the caller. So send
      // a message to self to actually do the shutdown.
      self.send(Shutdown)

    case Shutdown =>
      stopping.set(true)
      new Thread("CoarseGrainedExecutorBackend-stop-executor") {
        override def run(): Unit = {
          // executor.stop() will call `SparkEnv.stop()` which waits until RpcEnv stops totally.
          // However, if `executor.stop()` runs in some thread of RpcEnv, RpcEnv won't be able to
          // stop until `executor.stop()` returns, which becomes a dead-lock (See SPARK-14180).
          // Therefore, we put this line in a new thread.
          executor.stop()
        }
      }.start()
  }

　　　　　　

最后一个类SparkDeploySchedulerBackend（start）：

  var driverEndpoint: RpcEndpointRef = null

  protected def minRegisteredRatio: Double = _minRegisteredRatio

  override def start() {
    val properties = new ArrayBuffer[(String, String)]
    for ((key, value) <- scheduler.sc.conf.getAll) {
      if (key.startsWith("spark.")) {
        properties += ((key, value))
      }
    }

    // TODO (prashant) send conf instead of properties
    driverEndpoint = createDriverEndpointRef(properties)
  }

  protected def createDriverEndpointRef(
      properties: ArrayBuffer[(String, String)]): RpcEndpointRef = {
    rpcEnv.setupEndpoint(ENDPOINT_NAME, createDriverEndpoint(properties))
  }

  protected def createDriverEndpoint(properties: Seq[(String, String)]): DriverEndpoint = {
    new DriverEndpoint(rpcEnv, properties)
  }

　　　　　　

　　　　2.4调度模块：

　　2.4.1作业调度简介

　　　　　　DAGScheduler: 根据Job构建基于Stage的DAG（Directed Acyclic Graph有向无环图)，并提交Stage给TASkScheduler。 其划分Stage的依据是RDD之间的依赖的关系找出开销最小的调度方法，如下图

　　　　　　

　　　　　　

　　　　　　注：从最后一个Stage开始倒推,如果有依赖关系 就先解决父节点,如果没有依赖关系 就直接运行;这里做了一个简单的实验：Spark DAGSheduler生成Stage过程分析实验

　　　　　　2.4.2 任务调度简介：

　　　　　　TaskSchedulter: 将TaskSet提交给worker运行，每个Executor运行什么Task就是在此处分配的. TaskScheduler维护所有TaskSet，当Executor向Driver发生心跳时，TaskScheduler会根据资源剩余情况分配相应的Task。

　　　　　　另外TaskScheduler还维护着所有Task的运行标签，重试失败的Task。下图展示了TaskScheduler的作用

　　　　　　

　　　　　在不同运行模式中任务调度器具体为：

1. 1. 　　Spark on Standalone模式为TaskScheduler
   2. 　　YARN-Client模式为YarnClientClusterScheduler
   3. 　　YARN-Cluster模式为YarnClusterScheduler

 

　　3.运行模式

　　 3.1 standalone模式

---

• Standalone模式使用Spark自带的资源调度框架
• 采用Master/Slaves的典型架构，选用ZooKeeper来实现Master的HA
• 框架结构图如下:
• 
• 该模式主要的节点有Client节点、Master节点和Worker节点。其中Driver既可以运行在Master节点上中，也可以运行在本地Client端。当用spark-shell交互式工具提交Spark的Job时，Driver在Master节点上运行；当使用spark-submit工具提交Job或者在Eclips、IDEA等开发平台上使用”new SparkConf.setManager(“spark://master:7077”)”方式运行Spark任务时，Driver是运行在本地Client端上的
• 运行过程如下图：（参考至）
• 

1. SparkContext连接到Master，向Master注册并申请资源（CPU Core 和Memory）
2. Master根据SparkContext的资源申请要求和Worker心跳周期内报告的信息决定在哪个Worker上分配资源，然后在该Worker上获取资源，然后启动StandaloneExecutorBackend；
3. StandaloneExecutorBackend向SparkContext注册；
4. SparkContext将Applicaiton代码发送给StandaloneExecutorBackend；并且SparkContext解析Applicaiton代码，构建DAG图，并提交给DAG Scheduler分解成Stage（当碰到Action操作时，就会催生Job；每个Job中含有1个或多个Stage，Stage一般在获取外部数据和shuffle之前产生），然后以Stage（或者称为TaskSet）提交给Task Scheduler，Task Scheduler负责将Task分配到相应的Worker，最后提交给StandaloneExecutorBackend执行；
5. StandaloneExecutorBackend会建立Executor线程池，开始执行Task，并向SparkContext报告，直至Task完成
6. 所有Task完成后，SparkContext向Master注销，释放资源

　　

4 RDD实战

　　　　以下面一个按 A-Z 首字母分类，查找相同首字母下不同姓名总个数的例子来看一下 RDD 是如何运行起来的。

　　　　

sc.makeRDD(Seq("arachis","tony","lily","tom")).map{
      name => (name.charAt(0),name)
    }.groupByKey().mapValues{
      names => names.toSet.size //unique and count
    }.collect()

　

　提交Job collect

划分Stage

　提交Stage , 开始Task 运行调度

Stage0的DAG图，makeRDD => map ; 相应生成了两个RDD：ParallelCollectionRDD,MapPartitionsRDD

　Stage1 的DAG图，groupByKey => mapValues; 相应生成两个RDD：ShuffledRDD, MapPartitionsRDD

• 将这些术语串起来的运行层次图如下：
• 
• Job=多个stage，Stage=多个同种task, Task分为ShuffleMapTask和ResultTask，Dependency分为ShuffleDependency和NarrowDependency

链接：

　　Spark官网：http://spark.apache.org/docs/latest/cluster-overview.html

　　http://www.cnblogs.com/tgzhu/p/5818374.html