spark源码yarn-cluster模式任务提交的操作方法

这篇文章给大家介绍spark源码yarn-cluster模式任务提交的操作方法，内容非常详细，感兴趣的小伙伴们可以参考借鉴，希望对大家能有所帮助。

一，运行命令

bin/spark-submit \
--master yarn \
--deploy-mode cluster \
--class org.apache.spark.examples.SparkPi \
examples/jars/spark-examples_2.11-2.3.1.3.0.1.0-187.jar

二，任务提交流程图

三，启动脚本

查看spark-submit 脚本文件，程序入口为

exec "${SPARK_HOME}"/bin/spark-class org.apache.spark.deploy.SparkSubmit "$@“

查看${SPARK_HOME}"/bin/spark-class可知该脚本执行了java -cp main-class 命令启动了一个java进程,进程名为SparkSubmit，main函数在主类org.apache.spark.deploy.SparkSubmit中。

实际执行的具体命令为：

/etc/alternatives/jre/bin/java -Dhdp.version=3.0.1.0-187 -cp /usr/hdp/3.0.1.0-187/spark2/conf/:/usr/hdp/3.0.1.0-187/spark2/jars/*:/usr/hdp/3.0.1.0-187/hadoop/conf/ -Xmx1g org.apache.spark.deploy.SparkSubmit --master yarn --class org.apache.spark.examples.SparkPi examples/jars/spark-examples_2.11-2.3.1.3.0.1.0-187.jar

四，程序入口类org.apache.spark.deploy.SparkSubmit

该类有个伴生对象，其中有main函数,创建了SparkSubmit对象并执行doSubmit()；

override def main(args: Array[String]): Unit = {
  val submit = new SparkSubmit() {...}
  submit.doSubmit(args)
}

doSubmit 解析args参数，封装到appArgs:SparkSubmitArguments对象中，然后执行submit(appArgs, uninitLog)。

def doSubmit(args: Array[String]): Unit = {
  // Initialize logging if it hasn't been done yet. Keep track of whether logging needs to
  // be reset before the application starts.
  val uninitLog = initializeLogIfNecessary(true, silent = true)
  val appArgs = parseArguments(args)
  if (appArgs.verbose) {
    logInfo(appArgs.toString)
  }
  appArgs.action match {
    case SparkSubmitAction.SUBMIT => submit(appArgs, uninitLog)
    case SparkSubmitAction.KILL => kill(appArgs)
    case SparkSubmitAction.REQUEST_STATUS => requestStatus(appArgs)
    case SparkSubmitAction.PRINT_VERSION => printVersion()
  }
}

submit(appArgs, uninitLog) 调用 runMain(args: SparkSubmitArguments, uninitLog: Boolean)

private def runMain(args: SparkSubmitArguments, uninitLog: Boolean): Unit = {
  val (childArgs, childClasspath, sparkConf, childMainClass) = prepareSubmitEnvironment(args)
    .
    .
    .
  try {
     mainClass = Utils.classForName(childMainClass)
  } catch {...}
  val app: SparkApplication = if (classOf[SparkApplication].isAssignableFrom(mainClass)) {
    mainClass.getConstructor().newInstance().asInstanceOf[SparkApplication]
  } else {
    new JavaMainApplication(mainClass)
  }
    .
    .
    .
  try {
    app.start(childArgs.toArray, sparkConf)
  } catch {
    case t: Throwable =>
      throw findCause(t)
  }
}

这里mainClass十分重要，先判读mainClass是否是SparkApplication的子类，如果是则通过反射调用其构造器创建对象；

如果不是则创建一个JavaMainApplication（是SparkApplication的子类）对象并在其override def start(args: Array[String], conf: SparkConf)函数中利用反射执行mainClass中main函数。

SparkApplication创建完毕后执行其start(childArgs.toArray, sparkConf) 方法。

/**
 * Entry point for a Spark application. Implementations must provide a no-argument constructor.
 */
private[spark] trait SparkApplication {
  def start(args: Array[String], conf: SparkConf): Unit
}
/**
 * Implementation of SparkApplication that wraps a standard Java class with a "main" method.
 *
 * Configuration is propagated to the application via system properties, so running multiple
 * of these in the same JVM may lead to undefined behavior due to configuration leaks.
 */
private[deploy] class JavaMainApplication(klass: Class[_]) extends SparkApplication {
  override def start(args: Array[String], conf: SparkConf): Unit = {
    val mainMethod = klass.getMethod("main", new Array[String](0).getClass)
    if (!Modifier.isStatic(mainMethod.getModifiers)) {
      throw new IllegalStateException("The main method in the given main class must be static")
    }
    val sysProps = conf.getAll.toMap
    sysProps.foreach { case (k, v) =>
      sys.props(k) = v
    }
    mainMethod.invoke(null, args)
  }
}

如果**–deploy-mode** 是client mainClass的值由命令行参数 –class 决定，也就是org.apache.spark.examples.SparkPi。

这种情况下会在当前虚拟机中执行客户端代码，如果是其它条件情况会比较复杂。

以上文指定的运行命令为例，这里mainClass是org.apache.spark.deploy.yarn.YarnClusterApplication类class对象。

private[deploy] val YARN_CLUSTER_SUBMIT_CLASS =
  "org.apache.spark.deploy.yarn.YarnClusterApplication"
...
if (isYarnCluster) {
  childMainClass = YARN_CLUSTER_SUBMIT_CLASS
  if (args.isPython) {
    childArgs += ("--primary-py-file", args.primaryResource)
    childArgs += ("--class", "org.apache.spark.deploy.PythonRunner")
  } else if (args.isR) {
    val mainFile = new Path(args.primaryResource).getName
    childArgs += ("--primary-r-file", mainFile)
    childArgs += ("--class", "org.apache.spark.deploy.RRunner")
  } else {
    if (args.primaryResource != SparkLauncher.NO_RESOURCE) {
      childArgs += ("--jar", args.primaryResource)
    }
    childArgs += ("--class", args.mainClass)
  }
  if (args.childArgs != null) {
    args.childArgs.foreach { arg => childArgs += ("--arg", arg) }
  }
}

五，org.apache.spark.deploy.yarn.YarnClusterApplication类

该类在spark-yarn包中。

<dependency>
    <groupId>org.apache.spark</groupId>
    <artifactId>spark-yarn_${scala.version}</artifactId>
    <version>${spark.version}</version>
</dependency>

开始执行其override def start(args: Array[String], conf: SparkConf) 方法。

private[spark] class YarnClusterApplication extends SparkApplication {

  override def start(args: Array[String], conf: SparkConf): Unit = {
    // SparkSubmit would use yarn cache to distribute files & jars in yarn mode,
    // so remove them from sparkConf here for yarn mode.
    conf.remove(JARS)
    conf.remove(FILES)
    new Client(new ClientArguments(args), conf, null).run()
  }
}

SparkSubmi进程中创建一个客户端Client，该类是一个代理类其中包括YarnClient，执行run() 方法。

提交Application给yarn集群ResourceManager，提交成功后返回appid，

如果spark.submit.deployMode=cluster&&spark.yarn.submit.waitAppCompletion=true，

SparkSubmit进程会定期输出appId日志直到任务结束(monitorApplication(appId))，否则会输出一次日志然后退出。

def run(): Unit = {
    this.appId = submitApplication()
    if (!launcherBackend.isConnected() && fireAndForget) {
      val report = getApplicationReport(appId)
      val state = report.getYarnApplicationState
      logInfo(s"Application report for $appId (state: $state)")
      logInfo(formatReportDetails(report))
      if (state == YarnApplicationState.FAILED || state == YarnApplicationState.KILLED) {
        throw new SparkException(s"Application $appId finished with status: $state")
      }
    } else {
      val YarnAppReport(appState, finalState, diags) = monitorApplication(appId)
      if (appState == YarnApplicationState.FAILED || finalState == FinalApplicationStatus.FAILED) {
        diags.foreach { err =>
          logError(s"Application diagnostics message: $err")
        }
        throw new SparkException(s"Application $appId finished with failed status")
      }
      if (appState == YarnApplicationState.KILLED || finalState == FinalApplicationStatus.KILLED) {
        throw new SparkException(s"Application $appId is killed")
      }
      if (finalState == FinalApplicationStatus.UNDEFINED) {
        throw new SparkException(s"The final status of application $appId is undefined")
      }
    }
  }

继续跟踪submitApplication()

def submitApplication(): ApplicationId = {
    ResourceRequestHelper.validateResources(sparkConf)
    var appId: ApplicationId = null
    try {
      launcherBackend.connect()
      yarnClient.init(hadoopConf)
      yarnClient.start()
      logInfo("Requesting a new application from cluster with %d NodeManagers"
        .format(yarnClient.getYarnClusterMetrics.getNumNodeManagers))
      // Get a new application from our RM
      val newApp = yarnClient.createApplication()
      val newAppResponse = newApp.getNewApplicationResponse()
      appId = newAppResponse.getApplicationId()
      // The app staging dir based on the STAGING_DIR configuration if configured
      // otherwise based on the users home directory.
      val appStagingBaseDir = sparkConf.get(STAGING_DIR)
        .map { new Path(_, UserGroupInformation.getCurrentUser.getShortUserName) }
        .getOrElse(FileSystem.get(hadoopConf).getHomeDirectory())
      stagingDirPath = new Path(appStagingBaseDir, getAppStagingDir(appId))
      new CallerContext("CLIENT", sparkConf.get(APP_CALLER_CONTEXT),
        Option(appId.toString)).setCurrentContext()
      // Verify whether the cluster has enough resources for our AM
      verifyClusterResources(newAppResponse)
      // Set up the appropriate contexts to launch our AM
      val containerContext = createContainerLaunchContext(newAppResponse)
      val appContext = createApplicationSubmissionContext(newApp, containerContext)
      // Finally, submit and monitor the application
      logInfo(s"Submitting application $appId to ResourceManager")
      yarnClient.submitApplication(appContext)
      launcherBackend.setAppId(appId.toString)
      reportLauncherState(SparkAppHandle.State.SUBMITTED)
      appId
    } catch {
      case e: Throwable =>
        if (stagingDirPath != null) {
          cleanupStagingDir()
        }
        throw e
    }

该方法做了如下工作（对应于任务提交流程图中的1，2，3）：
1，向ResourceManager发送请求创建Application，获取全局唯一的
appId。
2，根据配置的缓存目录信息+appId信息，创建运行Application运行的缓存目录stagingDirPath。
3，verifyClusterResources 验证集群中是否有足够资源可用，没有的话抛出异常。
4，createContainerLaunchContext 创建Container，其中封装了Container进程的启动命令。
5，提交appContext。

查看createContainerLaunchContext(newAppResponse) 代码。

val amClass =
      if (isClusterMode) {
        Utils.classForName("org.apache.spark.deploy.yarn.ApplicationMaster").getName
      } else {
        Utils.classForName("org.apache.spark.deploy.yarn.ExecutorLauncher").getName
      }
...
// Command for the ApplicationMaster
    val commands = prefixEnv ++
      Seq(Environment.JAVA_HOME.$$() + "/bin/java", "-server") ++
      javaOpts ++ amArgs ++
      Seq(
        "1>", ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stdout",
        "2>", ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stderr")
    // TODO: it would be nicer to just make sure there are no null commands here
    val printableCommands = commands.map(s => if (s == null) "null" else s).toList
    amContainer.setCommands(printableCommands.asJava)

Container的启动代码大概为
bin/java -server org.apache.spark.deploy.yarn.ApplicationMaster --class …

六， org.apache.spark.deploy.yarn.ApplicationMaster 类。

yarn集群某一个NodeManager收到ResourceManager的命令，启动ApplicationMaster进程，对应任务提交流程图中的步骤4.
查看ApplicationMaster 伴生对象中的main方法。

def main(args: Array[String]): Unit = {
    SignalUtils.registerLogger(log)
    val amArgs = new ApplicationMasterArguments(args)
    val sparkConf = new SparkConf()
    if (amArgs.propertiesFile != null) {
      Utils.getPropertiesFromFile(amArgs.propertiesFile).foreach { case (k, v) =>
        sparkConf.set(k, v)
      }
    }
    // Set system properties for each config entry. This covers two use cases:
    // - The default configuration stored by the SparkHadoopUtil class
    // - The user application creating a new SparkConf in cluster mode
    //
    // Both cases create a new SparkConf object which reads these configs from system properties.
    sparkConf.getAll.foreach { case (k, v) =>
      sys.props(k) = v
    }
    val yarnConf = new YarnConfiguration(SparkHadoopUtil.newConfiguration(sparkConf))
    master = new ApplicationMaster(amArgs, sparkConf, yarnConf)
    val ugi = sparkConf.get(PRINCIPAL) match {
      // We only need to log in with the keytab in cluster mode. In client mode, the driver
      // handles the user keytab.
      case Some(principal) if master.isClusterMode =>
        val originalCreds = UserGroupInformation.getCurrentUser().getCredentials()
        SparkHadoopUtil.get.loginUserFromKeytab(principal, sparkConf.get(KEYTAB).orNull)
        val newUGI = UserGroupInformation.getCurrentUser()
       if (master.appAttemptId == null || master.appAttemptId.getAttemptId > 1) {
          // Re-obtain delegation tokens if this is not a first attempt, as they might be outdated
          // as of now. Add the fresh tokens on top of the original user's credentials (overwrite).
          // Set the context class loader so that the token manager has access to jars
          // distributed by the user.
          Utils.withContextClassLoader(master.userClassLoader) {
            val credentialManager = new HadoopDelegationTokenManager(sparkConf, yarnConf, null)
            credentialManager.obtainDelegationTokens(originalCreds)
          }
        }
        // Transfer the original user's tokens to the new user, since it may contain needed tokens
        // (such as those user to connect to YARN).
        newUGI.addCredentials(originalCreds)
        newUGI
      case _ =>
        SparkHadoopUtil.get.createSparkUser()
    }
    ugi.doAs(new PrivilegedExceptionAction[Unit]() {
      override def run(): Unit = System.exit(master.run())
    })
  }

创建了ApplicationMaster对象并执行其run() 方法。

 final def run(): Int = {
    try {
      val attemptID = if (isClusterMode) {
        // Set the web ui port to be ephemeral for yarn so we don't conflict with
        // other spark processes running on the same box
        System.setProperty(UI_PORT.key, "0")
        // Set the master and deploy mode property to match the requested mode.
        System.setProperty("spark.master", "yarn")
        System.setProperty(SUBMIT_DEPLOY_MODE.key, "cluster")
        // Set this internal configuration if it is running on cluster mode, this
        // configuration will be checked in SparkContext to avoid misuse of yarn cluster mode.
        System.setProperty("spark.yarn.app.id", appAttemptId.getApplicationId().toString())
        Option(appAttemptId.getAttemptId.toString)
      } else {
        None
      }
      new CallerContext(
        "APPMASTER", sparkConf.get(APP_CALLER_CONTEXT),
        Option(appAttemptId.getApplicationId.toString), attemptID).setCurrentContext()
      logInfo("ApplicationAttemptId: " + appAttemptId)
      // This shutdown hook should run *after* the SparkContext is shut down.
      val priority = ShutdownHookManager.SPARK_CONTEXT_SHUTDOWN_PRIORITY - 1
      ShutdownHookManager.addShutdownHook(priority) { () =>
        val maxAppAttempts = client.getMaxRegAttempts(sparkConf, yarnConf)
        val isLastAttempt = appAttemptId.getAttemptId() >= maxAppAttempts
        if (!finished) {
          // The default state of ApplicationMaster is failed if it is invoked by shut down hook.
          // This behavior is different compared to 1.x version.
          // If user application is exited ahead of time by calling System.exit(N), here mark
          // this application as failed with EXIT_EARLY. For a good shutdown, user shouldn't call
          // System.exit(0) to terminate the application.
          finish(finalStatus,
            ApplicationMaster.EXIT_EARLY,
            "Shutdown hook called before final status was reported.")
        }
        if (!unregistered) {
          // we only want to unregister if we don't want the RM to retry
          if (finalStatus == FinalApplicationStatus.SUCCEEDED || isLastAttempt) {
            unregister(finalStatus, finalMsg)
            cleanupStagingDir(new Path(System.getenv("SPARK_YARN_STAGING_DIR")))
          }
        }
      }
      if (isClusterMode) {
        runDriver()
      } else {
        runExecutorLauncher()
      }
    } catch {
      case e: Exception =>
        // catch everything else if not specifically handled
        logError("Uncaught exception: ", e)
        finish(FinalApplicationStatus.FAILED,
          ApplicationMaster.EXIT_UNCAUGHT_EXCEPTION,
          "Uncaught exception: " + StringUtils.stringifyException(e))
    } finally {
      try {
        metricsSystem.foreach { ms =>
          ms.report()
          ms.stop()
        }
      } catch {
        case e: Exception =>
          logWarning("Exception during stopping of the metric system: ", e)
      }
    }

    exitCode
  }

执行runDriver()方法。
userClassThread = startUserApplication() 启动了一个名为Driver的线程，该线程中通过反射执行命令行中**–class指定的类（org.apache.spark.examples.SparkPi）中的main**函数，初始化SparkContext。主线程唤醒后，向ResourceManager注册ApplicationMaster，步骤5；

private def runDriver(): Unit = {
    addAmIpFilter(None, System.getenv(ApplicationConstants.APPLICATION_WEB_PROXY_BASE_ENV))
    userClassThread = startUserApplication()
    // This a bit hacky, but we need to wait until the spark.driver.port property has
    // been set by the Thread executing the user class.
    logInfo("Waiting for spark context initialization...")
    val totalWaitTime = sparkConf.get(AM_MAX_WAIT_TIME)
    try {
      val sc = ThreadUtils.awaitResult(sparkContextPromise.future,
        Duration(totalWaitTime, TimeUnit.MILLISECONDS))
      if (sc != null) {
        val rpcEnv = sc.env.rpcEnv
        val userConf = sc.getConf
        val host = userConf.get(DRIVER_HOST_ADDRESS)
        val port = userConf.get(DRIVER_PORT)
        registerAM(host, port, userConf, sc.ui.map(_.webUrl), appAttemptId)
        val driverRef = rpcEnv.setupEndpointRef(
          RpcAddress(host, port),
          YarnSchedulerBackend.ENDPOINT_NAME)
        createAllocator(driverRef, userConf, rpcEnv, appAttemptId, distCacheConf)
      } else {
        // Sanity check; should never happen in normal operation, since sc should only be null
        // if the user app did not create a SparkContext.
        throw new IllegalStateException("User did not initialize spark context!")
      }
      resumeDriver()
      userClassThread.join()
    } catch {
      case e: SparkException if e.getCause().isInstanceOf[TimeoutException] =>
        logError(
          s"SparkContext did not initialize after waiting for $totalWaitTime ms. " +
           "Please check earlier log output for errors. Failing the application.")
        finish(FinalApplicationStatus.FAILED,
          ApplicationMaster.EXIT_SC_NOT_INITED,
          "Timed out waiting for SparkContext.")
    } finally {
      resumeDriver()
    }
  }

private def startUserApplication(): Thread = {
    logInfo("Starting the user application in a separate Thread")
    var userArgs = args.userArgs
    if (args.primaryPyFile != null && args.primaryPyFile.endsWith(".py")) {
      // When running pyspark, the app is run using PythonRunner. The second argument is the list
      // of files to add to PYTHONPATH, which Client.scala already handles, so it's empty.
      userArgs = Seq(args.primaryPyFile, "") ++ userArgs
    }
    if (args.primaryRFile != null &&
        (args.primaryRFile.endsWith(".R") || args.primaryRFile.endsWith(".r"))) {
      // TODO(davies): add R dependencies here
    }
    val mainMethod = userClassLoader.loadClass(args.userClass)
      .getMethod("main", classOf[Array[String]])
    val userThread = new Thread {
      override def run(): Unit = {
        try {
          if (!Modifier.isStatic(mainMethod.getModifiers)) {
            logError(s"Could not find static main method in object ${args.userClass}")
            finish(FinalApplicationStatus.FAILED, ApplicationMaster.EXIT_EXCEPTION_USER_CLASS)
          } else {
            mainMethod.invoke(null, userArgs.toArray)
            finish(FinalApplicationStatus.SUCCEEDED, ApplicationMaster.EXIT_SUCCESS)
            logDebug("Done running user class")
          }
        } catch {
          case e: InvocationTargetException =>
            e.getCause match {
              case _: InterruptedException =>
                // Reporter thread can interrupt to stop user class
              case SparkUserAppException(exitCode) =>
                val msg = s"User application exited with status $exitCode"
                logError(msg)
                finish(FinalApplicationStatus.FAILED, exitCode, msg)
              case cause: Throwable =>
                logError("User class threw exception: " + cause, cause)
                finish(FinalApplicationStatus.FAILED,
                  ApplicationMaster.EXIT_EXCEPTION_USER_CLASS,
                  "User class threw exception: " + StringUtils.stringifyException(cause))
            }
            sparkContextPromise.tryFailure(e.getCause())
        } finally {
          // Notify the thread waiting for the SparkContext, in case the application did not
          // instantiate one. This will do nothing when the user code instantiates a SparkContext
          // (with the correct master), or when the user code throws an exception (due to the
          // tryFailure above).
          sparkContextPromise.trySuccess(null)
        }
      }
    }
    userThread.setContextClassLoader(userClassLoader)
    userThread.setName("Driver")
    userThread.start()
    userThread
  }

注册完成后，主线程处理yarn返回的资源createAllocator(driverRef, userConf, rpcEnv, appAttemptId, distCacheConf)。

 private def createAllocator(
      driverRef: RpcEndpointRef,
      _sparkConf: SparkConf,
      rpcEnv: RpcEnv,
      appAttemptId: ApplicationAttemptId,
      distCacheConf: SparkConf): Unit = {
    // In client mode, the AM may be restarting after delegation tokens have reached their TTL. So
    // always contact the driver to get the current set of valid tokens, so that local resources can
    // be initialized below.
    if (!isClusterMode) {
      val tokens = driverRef.askSync[Array[Byte]](RetrieveDelegationTokens)
      if (tokens != null) {
        SparkHadoopUtil.get.addDelegationTokens(tokens, _sparkConf)
      }
    }
    val appId = appAttemptId.getApplicationId().toString()
    val driverUrl = RpcEndpointAddress(driverRef.address.host, driverRef.address.port,
      CoarseGrainedSchedulerBackend.ENDPOINT_NAME).toString
    val localResources = prepareLocalResources(distCacheConf)
    // Before we initialize the allocator, let's log the information about how executors will
    // be run up front, to avoid printing this out for every single executor being launched.
    // Use placeholders for information that changes such as executor IDs.
    logInfo {
      val executorMemory = _sparkConf.get(EXECUTOR_MEMORY).toInt
      val executorCores = _sparkConf.get(EXECUTOR_CORES)
      val dummyRunner = new ExecutorRunnable(None, yarnConf, _sparkConf, driverUrl, "<executorId>",
        "<hostname>", executorMemory, executorCores, appId, securityMgr, localResources,
        ResourceProfile.DEFAULT_RESOURCE_PROFILE_ID)
      dummyRunner.launchContextDebugInfo()
    }
    allocator = client.createAllocator(
      yarnConf,
      _sparkConf,
      appAttemptId,
      driverUrl,
      driverRef,
      securityMgr,
      localResources)
    // Initialize the AM endpoint *after* the allocator has been initialized. This ensures
    // that when the driver sends an initial executor request (e.g. after an AM restart),
    // the allocator is ready to service requests.
    rpcEnv.setupEndpoint("YarnAM", new AMEndpoint(rpcEnv, driverRef))
    allocator.allocateResources()
    val ms = MetricsSystem.createMetricsSystem(MetricsSystemInstances.APPLICATION_MASTER,
      sparkConf, securityMgr)
    val prefix = _sparkConf.get(YARN_METRICS_NAMESPACE).getOrElse(appId)
    ms.registerSource(new ApplicationMasterSource(prefix, allocator))
    // do not register static sources in this case as per SPARK-25277
    ms.start(false)
    metricsSystem = Some(ms)
    reporterThread = launchReporterThread()
  }

只看关键代码allocator.allocateResources()，处理分配的资源。

def allocateResources(): Unit = synchronized {
    updateResourceRequests()
    val progressIndicator = 0.1f
    // Poll the ResourceManager. This doubles as a heartbeat if there are no pending container
    // requests.
    val allocateResponse = amClient.allocate(progressIndicator)
    val allocatedContainers = allocateResponse.getAllocatedContainers()
    allocatorBlacklistTracker.setNumClusterNodes(allocateResponse.getNumClusterNodes)
    if (allocatedContainers.size > 0) {
      logDebug(("Allocated containers: %d. Current executor count: %d. " +
        "Launching executor count: %d. Cluster resources: %s.")
        .format(
          allocatedContainers.size,
          runningExecutors.size,
          numExecutorsStarting.get,
          allocateResponse.getAvailableResources))
      handleAllocatedContainers(allocatedContainers.asScala)
    }
   val completedContainers = allocateResponse.getCompletedContainersStatuses()
    if (completedContainers.size > 0) {
      logDebug("Completed %d containers".format(completedContainers.size))
      processCompletedContainers(completedContainers.asScala)
      logDebug("Finished processing %d completed containers. Current running executor count: %d."
        .format(completedContainers.size, runningExecutors.size))
    }
  }

如果分配的Container数量大于0，调用** handleAllocatedContainers(allocatedContainers.asScala)**

def handleAllocatedContainers(allocatedContainers: Seq[Container]): Unit = {
    val containersToUse = new ArrayBuffer[Container](allocatedContainers.size)
    // Match incoming requests by host
    val remainingAfterHostMatches = new ArrayBuffer[Container]
    for (allocatedContainer <- allocatedContainers) {
      matchContainerToRequest(allocatedContainer, allocatedContainer.getNodeId.getHost,
        containersToUse, remainingAfterHostMatches)
    }
    // Match remaining by rack. Because YARN's RackResolver swallows thread interrupts
    // (see SPARK-27094), which can cause this code to miss interrupts from the AM, use
    // a separate thread to perform the operation.
    val remainingAfterRackMatches = new ArrayBuffer[Container]
    if (remainingAfterHostMatches.nonEmpty) {
      var exception: Option[Throwable] = None
      val thread = new Thread("spark-rack-resolver") {
        override def run(): Unit = {
          try {
            for (allocatedContainer <- remainingAfterHostMatches) {
              val rack = resolver.resolve(allocatedContainer.getNodeId.getHost)
              matchContainerToRequest(allocatedContainer, rack, containersToUse,
                remainingAfterRackMatches)
            }
          } catch {
            case e: Throwable =>
              exception = Some(e)
          }
        }
      }
      thread.setDaemon(true)
      thread.start()
      try {
        thread.join()
      } catch {
        case e: InterruptedException =>
          thread.interrupt()
          throw e
      }
      if (exception.isDefined) {
        throw exception.get
      }
    }
    // Assign remaining that are neither node-local nor rack-local
    val remainingAfterOffRackMatches = new ArrayBuffer[Container]
    for (allocatedContainer <- remainingAfterRackMatches) {
      matchContainerToRequest(allocatedContainer, ANY_HOST, containersToUse,
        remainingAfterOffRackMatches)
    }
    if (remainingAfterOffRackMatches.nonEmpty) {
      logDebug(s"Releasing ${remainingAfterOffRackMatches.size} unneeded containers that were " +
        s"allocated to us")
      for (container <- remainingAfterOffRackMatches) {
        internalReleaseContainer(container)
      }
    }
    runAllocatedContainers(containersToUse)
    logInfo("Received %d containers from YARN, launching executors on %d of them."
      .format(allocatedContainers.size, containersToUse.size))
  }

这里会根据主机host，机架rack等信息队container进行分配。完成后启动Container，runAllocatedContainers(containersToUse)。

  private val launcherPool = ThreadUtils.newDaemonCachedThreadPool(
    "ContainerLauncher", sparkConf.get(CONTAINER_LAUNCH_MAX_THREADS))

创建线程池launcherPool。

  /**
   * Launches executors in the allocated containers.
   */
  private def runAllocatedContainers(containersToUse: ArrayBuffer[Container]): Unit = {
    for (container <- containersToUse) {
      executorIdCounter += 1
      val executorHostname = container.getNodeId.getHost
      val containerId = container.getId
      val executorId = executorIdCounter.toString
      assert(container.getResource.getMemory >= resource.getMemory)
      logInfo(s"Launching container $containerId on host $executorHostname " +
        s"for executor with ID $executorId")
      def updateInternalState(): Unit = synchronized {
        runningExecutors.add(executorId)
        numExecutorsStarting.decrementAndGet()
        executorIdToContainer(executorId) = container
        containerIdToExecutorId(container.getId) = executorId
        val containerSet = allocatedHostToContainersMap.getOrElseUpdate(executorHostname,
          new HashSet[ContainerId])
        containerSet += containerId
        allocatedContainerToHostMap.put(containerId, executorHostname)
      }
      if (runningExecutors.size() < targetNumExecutors) {
        numExecutorsStarting.incrementAndGet()
        if (launchContainers) {
          launcherPool.execute(() => {
            try {
              new ExecutorRunnable(
                Some(container),
                conf,
                sparkConf,
                driverUrl,
                executorId,
                executorHostname,
                executorMemory,
                executorCores,
                appAttemptId.getApplicationId.toString,
                securityMgr,
                localResources,
                ResourceProfile.DEFAULT_RESOURCE_PROFILE_ID // use until fully supported
              ).run()
              updateInternalState()
            } catch {
              case e: Throwable =>
                numExecutorsStarting.decrementAndGet()
                if (NonFatal(e)) {
                  logError(s"Failed to launch executor $executorId on container $containerId", e)
                  // Assigned container should be released immediately
                  // to avoid unnecessary resource occupation.
                  amClient.releaseAssignedContainer(containerId)
                } else {
                  throw e
                }
            }
          })
        } else {
          // For test only
          updateInternalState()
        }
      } else {
        logInfo(("Skip launching executorRunnable as running executors count: %d " +
          "reached target executors count: %d.").format(
          runningExecutors.size, targetNumExecutors))
      }
    }
  }

查看ExecutorRunnable 类，其中nmClient = NMClient.createNMClient()， NodeManager客户端，负责于NodeManager交互；其prepareCommand() 方法拼接了一个进程启动命令，大体格式为：

bin/java -server org.apache.spark.executor.YarnCoarseGrainedExecutorBackend ...

ApplicationMaster进程中的launcherPool线程池，会根据Container的个数挨个启动线程ExecutorRunnable，ExecutorRunnable中的NMClient会将拼接好的jvm启动命令发送给相关的NodeManager，启动Container进程，进程名为YarnCoarseGrainedExecutorBackend。
ExecutorRunnable完整代码：

private[yarn] class ExecutorRunnable(
    container: Option[Container],
    conf: YarnConfiguration,
    sparkConf: SparkConf,
    masterAddress: String,
    executorId: String,
    hostname: String,
    executorMemory: Int,
    executorCores: Int,
    appId: String,
    securityMgr: SecurityManager,
    localResources: Map[String, LocalResource],
    resourceProfileId: Int) extends Logging {
  var rpc: YarnRPC = YarnRPC.create(conf)
  var nmClient: NMClient = _
  def run(): Unit = {
    logDebug("Starting Executor Container")
    nmClient = NMClient.createNMClient()
    nmClient.init(conf)
    nmClient.start()
    startContainer()
  }
  def launchContextDebugInfo(): String = {
    val commands = prepareCommand()
    val env = prepareEnvironment()
    s"""
    |===============================================================================
    |Default YARN executor launch context:
    |  env:
    |${Utils.redact(sparkConf, env.toSeq).map { case (k, v) => s"    $k -> $v\n" }.mkString}
    |  command:
    |    ${Utils.redactCommandLineArgs(sparkConf, commands).mkString(" \\ \n      ")}
    |
    |  resources:
    |${localResources.map { case (k, v) => s"    $k -> $v\n" }.mkString}
    |===============================================================================""".stripMargin
  }
  def startContainer(): java.util.Map[String, ByteBuffer] = {
    val ctx = Records.newRecord(classOf[ContainerLaunchContext])
      .asInstanceOf[ContainerLaunchContext]
    val env = prepareEnvironment().asJava
    ctx.setLocalResources(localResources.asJava)
    ctx.setEnvironment(env)
    val credentials = UserGroupInformation.getCurrentUser().getCredentials()
    val dob = new DataOutputBuffer()
    credentials.writeTokenStorageToStream(dob)
    ctx.setTokens(ByteBuffer.wrap(dob.getData()))
    val commands = prepareCommand()
    ctx.setCommands(commands.asJava)
    ctx.setApplicationACLs(
      YarnSparkHadoopUtil.getApplicationAclsForYarn(securityMgr).asJava)
    // If external shuffle service is enabled, register with the Yarn shuffle service already
    // started on the NodeManager and, if authentication is enabled, provide it with our secret
    // key for fetching shuffle files later
    if (sparkConf.get(SHUFFLE_SERVICE_ENABLED)) {
      val secretString = securityMgr.getSecretKey()
      val secretBytes =
        if (secretString != null) {
          // This conversion must match how the YarnShuffleService decodes our secret
          JavaUtils.stringToBytes(secretString)
        } else {
          // Authentication is not enabled, so just provide dummy metadata
          ByteBuffer.allocate(0)
        }
      ctx.setServiceData(Collections.singletonMap("spark_shuffle", secretBytes))
    }
    // Send the start request to the ContainerManager
    try {
      nmClient.startContainer(container.get, ctx)
    } catch {
      case ex: Exception =>
        throw new SparkException(s"Exception while starting container ${container.get.getId}" +
          s" on host $hostname", ex)
    }
  }
  private def prepareCommand(): List[String] = {
    // Extra options for the JVM
    val javaOpts = ListBuffer[String]()
    // Set the JVM memory
    val executorMemoryString = executorMemory + "m"
    javaOpts += "-Xmx" + executorMemoryString
    // Set extra Java options for the executor, if defined
    sparkConf.get(EXECUTOR_JAVA_OPTIONS).foreach { opts =>
      val subsOpt = Utils.substituteAppNExecIds(opts, appId, executorId)
      javaOpts ++= Utils.splitCommandString(subsOpt).map(YarnSparkHadoopUtil.escapeForShell)
    }
    // Set the library path through a command prefix to append to the existing value of the
    // env variable.
    val prefixEnv = sparkConf.get(EXECUTOR_LIBRARY_PATH).map { libPath =>
      Client.createLibraryPathPrefix(libPath, sparkConf)
    }
    javaOpts += "-Djava.io.tmpdir=" +
      new Path(Environment.PWD.$$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR)
    // Certain configs need to be passed here because they are needed before the Executor
    // registers with the Scheduler and transfers the spark configs. Since the Executor backend
    // uses RPC to connect to the scheduler, the RPC settings are needed as well as the
    // authentication settings.
    sparkConf.getAll
      .filter { case (k, v) => SparkConf.isExecutorStartupConf(k) }
      .foreach { case (k, v) => javaOpts += YarnSparkHadoopUtil.escapeForShell(s"-D$k=$v") }
    // Commenting it out for now - so that people can refer to the properties if required. Remove
    // it once cpuset version is pushed out.
    // The context is, default gc for server class machines end up using all cores to do gc - hence
    // if there are multiple containers in same node, spark gc effects all other containers
    // performance (which can also be other spark containers)
    // Instead of using this, rely on cpusets by YARN to enforce spark behaves 'properly' in
    // multi-tenant environments. Not sure how default java gc behaves if it is limited to subset
    // of cores on a node.
    /*
        else {
          // If no java_opts specified, default to using -XX:+CMSIncrementalMode
          // It might be possible that other modes/config is being done in
          // spark.executor.extraJavaOptions, so we don't want to mess with it.
          // In our expts, using (default) throughput collector has severe perf ramifications in
          // multi-tenant machines
          // The options are based on
          // http://www.oracle.com/technetwork/java/gc-tuning-5-138395.html#0.0.0.%20When%20to%20Use
          // %20the%20Concurrent%20Low%20Pause%20Collector|outline
          javaOpts += "-XX:+UseConcMarkSweepGC"
          javaOpts += "-XX:+CMSIncrementalMode"
          javaOpts += "-XX:+CMSIncrementalPacing"
          javaOpts += "-XX:CMSIncrementalDutyCycleMin=0"
          javaOpts += "-XX:CMSIncrementalDutyCycle=10"
        }
    */
    // For log4j configuration to reference
    javaOpts += ("-Dspark.yarn.app.container.log.dir=" + ApplicationConstants.LOG_DIR_EXPANSION_VAR)
    val userClassPath = Client.getUserClasspath(sparkConf).flatMap { uri =>
      val absPath =
        if (new File(uri.getPath()).isAbsolute()) {
          Client.getClusterPath(sparkConf, uri.getPath())
        } else {
          Client.buildPath(Environment.PWD.$(), uri.getPath())
        }
      Seq("--user-class-path", "file:" + absPath)
    }.toSeq
    YarnSparkHadoopUtil.addOutOfMemoryErrorArgument(javaOpts)
    val commands = prefixEnv ++
      Seq(Environment.JAVA_HOME.$$() + "/bin/java", "-server") ++
      javaOpts ++
      Seq("org.apache.spark.executor.YarnCoarseGrainedExecutorBackend",
        "--driver-url", masterAddress,
        "--executor-id", executorId,
        "--hostname", hostname,
        "--cores", executorCores.toString,
        "--app-id", appId,
        "--resourceProfileId", resourceProfileId.toString) ++
      userClassPath ++
      Seq(
        s"1>${ApplicationConstants.LOG_DIR_EXPANSION_VAR}/stdout",
        s"2>${ApplicationConstants.LOG_DIR_EXPANSION_VAR}/stderr")
    // TODO: it would be nicer to just make sure there are no null commands here
    commands.map(s => if (s == null) "null" else s).toList
  }
  private def prepareEnvironment(): HashMap[String, String] = {
    val env = new HashMap[String, String]()
    Client.populateClasspath(null, conf, sparkConf, env, sparkConf.get(EXECUTOR_CLASS_PATH))
    System.getenv().asScala.filterKeys(_.startsWith("SPARK"))
      .foreach { case (k, v) => env(k) = v }
    sparkConf.getExecutorEnv.foreach { case (key, value) =>
      if (key == Environment.CLASSPATH.name()) {
        // If the key of env variable is CLASSPATH, we assume it is a path and append it.
        // This is kept for backward compatibility and consistency with hadoop
        YarnSparkHadoopUtil.addPathToEnvironment(env, key, value)
      } else {
        // For other env variables, simply overwrite the value.
        env(key) = value
      }
    }
   env
  }
}

关于spark源码yarn-cluster模式任务提交的操作方法就分享到这里了，希望以上内容可以对大家有一定的帮助，可以学到更多知识。如果觉得文章不错，可以把它分享出去让更多的人看到。