Spark AQE 导致的 Driver OOM问题

背景

最近在做Spark 3.1 升级 Spark 3.5的过程中，遇到了一批SQL在运行的过程中 Driver OOM的情况，排查到是AQE开启导致的问题，再次分析记录一下，顺便了解一下Spark中指标的事件处理情况

结论

SQLAppStatusListener 类在内存中存放着 一个整个SQL查询链的所有stage以及stage的指标信息，在AQE中 一个job会被拆分成很多job，甚至几百上千的job，这个时候 stageMetrics的数据就会成百上倍的被存储在内存中，从而导致Driver OOM。
解决方法：

1. 关闭AQE spark.sql.adaptive.enabled false
2. 合并对应的PR-SPARK-45439

分析

背景知识：对于一个完整链接的sql语句来说(比如说从 读取数据源，到 数据处理操作，再到插入hive表),这可以称其为一个最小的SQL执行单元，这最小的数据执行单元在Spark内部是可以跟踪的,也就是用executionId来进行跟踪的。
对于一个sql，举例来说 :

insert into  TableA select * from TableB;

在生成 物理计划的过程中会调用 QueryExecution.assertOptimized 方法，该方法会触发eagerlyExecuteCommands调用，最终会到SQLExecution.withNewExecutionId方法：

  def assertOptimized(): Unit = optimizedPlan

  ...
  lazy val commandExecuted: LogicalPlan = mode match {
    case CommandExecutionMode.NON_ROOT => analyzed.mapChildren(eagerlyExecuteCommands)
    case CommandExecutionMode.ALL => eagerlyExecuteCommands(analyzed)
    case CommandExecutionMode.SKIP => analyzed
  }
  ...
  lazy val optimizedPlan: LogicalPlan = {
  // We need to materialize the commandExecuted here because optimizedPlan is also tracked under
  // the optimizing phase
  assertCommandExecuted()
  executePhase(QueryPlanningTracker.OPTIMIZATION) {
    // clone the plan to avoid sharing the plan instance between different stages like analyzing,
    // optimizing and planning.
    val plan =
      sparkSession.sessionState.optimizer.executeAndTrack(withCachedData.clone(), tracker)
    // We do not want optimized plans to be re-analyzed as literals that have been constant
    // folded and such can cause issues during analysis. While `clone` should maintain the
    // `analyzed` state of the LogicalPlan, we set the plan as analyzed here as well out of
    // paranoia.
    plan.setAnalyzed()
    plan
  }
  
  def assertCommandExecuted(): Unit = commandExecuted
  ...
  private def eagerlyExecuteCommands(p: LogicalPlan) = p transformDown {
    case c: Command =>
      // Since Command execution will eagerly take place here,
      // and in most cases be the bulk of time and effort,
      // with the rest of processing of the root plan being just outputting command results,
      // for eagerly executed commands we mark this place as beginning of execution.
      tracker.setReadyForExecution()
      val qe = sparkSession.sessionState.executePlan(c, CommandExecutionMode.NON_ROOT)
      val name = commandExecutionName(c)
      val result = QueryExecution.withInternalError(s"Eagerly executed $name failed.") {
        SQLExecution.withNewExecutionId(qe, Some(name)) {
          qe.executedPlan.executeCollect()
        }
      }  

而SQLExecution.withNewExecutionId主要的作用是设置当前计划的所属的executionId:

    val executionId = SQLExecution.nextExecutionId
    sc.setLocalProperty(EXECUTION_ID_KEY, executionId.toString)

该EXECUTION_ID_KEY的值会在JobStart的时候传递给Event，以便记录跟踪整个执行过程中的指标信息。
同时我们在方法中eagerlyExecuteCommands看到qe.executedPlan.executeCollect()这是具体的执行方法，针对于insert into 操作来说，物理计划就是
InsertIntoHadoopFsRelationCommand，这里的run方法最终会流转到DAGScheduler.submitJob方法：

    eventProcessLoop.post(JobSubmitted(
      jobId, rdd, func2, partitions.toArray, callSite, waiter,
      JobArtifactSet.getActiveOrDefault(sc),
      Utils.cloneProperties(properties)))

最终会被DAGScheduler.handleJobSubmitted处理，其中会发送SparkListenerJobStart事件：

    listenerBus.post(
      SparkListenerJobStart(job.jobId, jobSubmissionTime, stageInfos,
        Utils.cloneProperties(properties)))

该事件会被SQLAppStatusListener捕获，从而转到onJobStart处理，这里有会涉及到指标信息的存储，这里我们截图出dump的内存占用情况：

可以看到 SQLAppStatusListener 的 LiveStageMetrics 占用很大，也就是 accumIdsToMetricType占用很大

那在AQE中是怎么回事呢？
我们知道再AQE中，任务会从source节点按照shuffle进行分割，从而形成单独的job，从而生成对应的shuffle指标，具体的分割以及执行代码在AdaptiveSparkPlanExec.getFinalPhysicalPlan中，如下:

      var result = createQueryStages(currentPhysicalPlan)
      val events = new LinkedBlockingQueue[StageMaterializationEvent]()
      val errors = new mutable.ArrayBuffer[Throwable]()
      var stagesToReplace = Seq.empty[QueryStageExec]
      while (!result.allChildStagesMaterialized) {
        currentPhysicalPlan = result.newPlan
        if (result.newStages.nonEmpty) {
          stagesToReplace = result.newStages ++ stagesToReplace
          executionId.foreach(onUpdatePlan(_, result.newStages.map(_.plan)))

          // SPARK-33933: we should submit tasks of broadcast stages first, to avoid waiting
          // for tasks to be scheduled and leading to broadcast timeout.
          // This partial fix only guarantees the start of materialization for BroadcastQueryStage
          // is prior to others, but because the submission of collect job for broadcasting is
          // running in another thread, the issue is not completely resolved.
          val reorderedNewStages = result.newStages
            .sortWith {
              case (_: BroadcastQueryStageExec, _: BroadcastQueryStageExec) => false
              case (_: BroadcastQueryStageExec, _) => true
              case _ => false
            }

          // Start materialization of all new stages and fail fast if any stages failed eagerly
          reorderedNewStages.foreach { stage =>
            try {
              stage.materialize().onComplete { res =>
                if (res.isSuccess) {
                  events.offer(StageSuccess(stage, res.get))
                } else {
                  events.offer(StageFailure(stage, res.failed.get))
                }
                // explicitly clean up the resources in this stage
                stage.cleanupResources()
              }(AdaptiveSparkPlanExec.executionContext)

这里就是得看stage.materialize()这个方法，这两个stage只有两类：BroadcastQueryStageExec 和 ShuffleQueryStageExec，
这两个物理计划稍微分析一下如下:

• BroadcastQueryStageExec
  数据流如下：

  broadcast.submitBroadcastJob
        ||
        \/
  promise.future
        ||
        \/
  relationFuture
        ||
        \/
  child.executeCollectIterator()
  
  

  其中 promise的设置在relationFuture方法中，而relationFuture 会被doPrepare调用,而submitBroadcastJob会调用executeQuery,从而调用doPrepare,executeCollectIterator()最终也会发送JobSubmitted事件，分析和上面的一样
• ShuffleQueryStageExec

   shuffle.submitShuffleJob
        ||
        \/
   sparkContext.submitMapStage(shuffleDependency)
        ||
        \/
   dagScheduler.submitMapStage
  
  

该submitMapStage会发送MapStageSubmitted事件:

    eventProcessLoop.post(MapStageSubmitted(
      jobId, dependency, callSite, waiter, JobArtifactSet.getActiveOrDefault(sc),
      Utils.cloneProperties(properties)))

最终会被DAGScheduler.handleMapStageSubmitted处理，其中会发送SparkListenerJobStart事件：

    listenerBus.post(
      SparkListenerJobStart(job.jobId, jobSubmissionTime, stageInfos,
        Utils.cloneProperties(properties)))

该事件会被SQLAppStatusListener捕获，从而转到onJobStart处理:

  private val liveExecutions = new ConcurrentHashMap[Long, LiveExecutionData]()
  private val stageMetrics = new ConcurrentHashMap[Int, LiveStageMetrics]()
   ...
 
  override def onJobStart(event: SparkListenerJobStart): Unit = {
    val executionIdString = event.properties.getProperty(SQLExecution.EXECUTION_ID_KEY)
    if (executionIdString == null) {
      // This is not a job created by SQL
      return
    }

    val executionId = executionIdString.toLong
    val jobId = event.jobId
    val exec = Option(liveExecutions.get(executionId))

该方法会获取事件中的executionId,在AQE中，同一个执行单元的executionId是一样的，所以stageMetrics内存占用会越来越大。
而这里指标的更新是在AdaptiveSparkPlanExec.onUpdatePlan等方法中。

这样整个事件的数据流以及问题的产生原因就应该很清楚了。

其他

为啥AQE以后多个Job还是共享一个executionId呢？因为原则上来说，如果没有开启AQE之前，一个SQL执行单元的是属于同一个Job的，开启了AQE之后，因为AQE的原因，一个Job被拆成了了多个Job，但是从逻辑上来说，还是属于同一个SQL处理单元的所以还是得归属到一次执行中。