Spark Streaming job的生成及数据清理总结

关于这次总结还是要从一个bug说起。。。。。。。

场景描述：项目的基本处理流程为：从文件系统读取每隔一分钟上传的日志并由Spark Streaming进行计算消费，最后将结果写入InfluxDB中，然后在监控系统中进行展示，监控。这里的spark版本为2.2.1。

Bug：程序开发完成之后，每个batch处理时间在15~20s左右，上线之后一直在跑，监控系统中数据也没有什么异常，sparkui中只关注了任务处理时间，其他并没有在意。后来程序运行了2天18个小时之后，监控系统发出报警NO DATA，先去数据库查数据，确实没有数据，在去sparkui看程序并没有结束，状态还是RUNNING，但是不处理任务，就在那里卡住了，后来看日志发现报了内存溢出异常：

Exception in thread "dag-scheduler-event-loop" java.lang.OutOfMemoryError: Java heap space
    at java.nio.HeapByteBuffer.<init>(HeapByteBuffer.java:57)
    at java.nio.ByteBuffer.allocate(ByteBuffer.java:335)
    at org.apache.spark.broadcast.TorrentBroadcast$$anonfun$3.apply(TorrentBroadcast.scala:271)
    at org.apache.spark.broadcast.TorrentBroadcast$$anonfun$3.apply(TorrentBroadcast.scala:271)
    at org.apache.spark.util.io.ChunkedByteBufferOutputStream.allocateNewChunkIfNeeded(ChunkedByteBufferOutputStream.scala:87)
    at org.apache.spark.util.io.ChunkedByteBufferOutputStream.write(ChunkedByteBufferOutputStream.scala:75)
    at net.jpountz.lz4.LZ4BlockOutputStream.flushBufferedData(LZ4BlockOutputStream.java:205)
    at net.jpountz.lz4.LZ4BlockOutputStream.finish(LZ4BlockOutputStream.java:235)
    at net.jpountz.lz4.LZ4BlockOutputStream.close(LZ4BlockOutputStream.java:175)
    at java.io.ObjectOutputStream$BlockDataOutputStream.close(ObjectOutputStream.java:1828)
    at java.io.ObjectOutputStream.close(ObjectOutputStream.java:742)
    at org.apache.spark.serializer.JavaSerializationStream.close(JavaSerializer.scala:57)
    at org.apache.spark.broadcast.TorrentBroadcast$$anonfun$blockifyObject$1.apply$mcV$sp(TorrentBroadcast.scala:278)
    at org.apache.spark.util.Utils$.tryWithSafeFinally(Utils.scala:1346)
    at org.apache.spark.broadcast.TorrentBroadcast$.blockifyObject(TorrentBroadcast.scala:277)
    at org.apache.spark.broadcast.TorrentBroadcast.writeBlocks(TorrentBroadcast.scala:126)
    at org.apache.spark.broadcast.TorrentBroadcast.<init>(TorrentBroadcast.scala:88)
    at org.apache.spark.broadcast.TorrentBroadcastFactory.newBroadcast(TorrentBroadcastFactory.scala:34)
    at org.apache.spark.broadcast.BroadcastManager.newBroadcast(BroadcastManager.scala:56)
    at org.apache.spark.SparkContext.broadcast(SparkContext.scala:1488)
    at org.apache.spark.scheduler.DAGScheduler.submitMissingTasks(DAGScheduler.scala:1006)
    at org.apache.spark.scheduler.DAGScheduler.org$apache$spark$scheduler$DAGScheduler$$submitStage(DAGScheduler.scala:930)
    at org.apache.spark.scheduler.DAGScheduler$$anonfun$submitWaitingChildStages$6.apply(DAGScheduler.scala:776)
    at org.apache.spark.scheduler.DAGScheduler$$anonfun$submitWaitingChildStages$6.apply(DAGScheduler.scala:775)
    at scala.collection.IndexedSeqOptimized$class.foreach(IndexedSeqOptimized.scala:33)
    at scala.collection.mutable.ArrayOps$ofRef.foreach(ArrayOps.scala:186)
    at org.apache.spark.scheduler.DAGScheduler.submitWaitingChildStages(DAGScheduler.scala:775)
    at org.apache.spark.scheduler.DAGScheduler.handleTaskCompletion(DAGScheduler.scala:1278)
    at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.doOnReceive(DAGScheduler.scala:1729)
    at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:1687)
    at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onReceive(DAGScheduler.scala:1676)
    at org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:48)

后来以为是程序里面的资源没有回收，仔细排查了一遍代码，也没找出来问题，后来在本地跑，发现sparkUI中EXECUTOR中storage memory和RDD blocks会一直增加，虽然每个Batch后俩者会下降，但是每一个Batch之后和上一个batch比较还是增加的。

解决：由于是Storage memory和RDD blocks在增长，觉得和内存相关，用内存调优改了下参数还是不行，然后以为是contentcleaner问题，把它调成5分钟一次也不行，后来在我的另一个电脑跑的时候发现没有问题了，这个电脑上的spark 版本是2.3.0；现在可以确定是版本问题，就直接去官网2.2.2版本里面找关于内存溢出修复的bug，当时就下载了2.2.2，然后以跑程序还是和原来2.2.1一样，再然后就心态崩了，也没想着去看2.3.0的BUG修复了，当时我在一个知识星球提问过这个问题，后来星球的主人帮我解决了这个问题，原来这个问题在2.3.0里面才被解决，具体网址：https://issues.apache.org/jira/browse/SPARK-21357，原因是因为FileInputSream会重写Dstream中的clearMetadata方法，但是在FileInputStream中claerMetadata方法只是清理了文件并没有清理generatedRDDs，因此才会出现内存溢出。

总结：本次bug本来在确定了版本问题之后，理应很好解决，但是由于自身原因，多走了弯路，后来得高人相助才得以顺利解决问题。由此也发现了自己的一些问题，遇到问题不能只能留在表面，要深入代码，在了解原理的基础上在了解具体实现，在遇到问题是才能快速定位问题，并找到解决办法。下面就是这次bug之后翻看spark streaming源码之后对出现这个bug的前因后果的分析。

bug分析：spark Streaming程序只要启动就会一直的运转，期间从数据源得到数据，然后消费，最后输出，在每一个的batch里面，都会根据具体的业务逻辑生成对应的jod，然后spark就处理提交的job，这里只要明白了job的生成及生成之后对缓存的数据的处理，也就好理解这个bug的出现原因了。

StreamingContent在启动之后，会启动JobScheduler；在JobSchedluer里面会启动JobGenerator和ReceiveTracker；JobGenerator负责job相关的处理，ReceiveTracker负责Receive分发和worker端的receive通信，并处理其发来的信息。

如下是JobSchedluer的start方法：

 def start(): Unit = synchronized {
    if (eventLoop != null) return // scheduler has already been started

    logDebug("Starting JobScheduler")
    eventLoop = new EventLoop[JobSchedulerEvent]("JobScheduler") {
      override protected def onReceive(event: JobSchedulerEvent): Unit = processEvent(event)

      override protected def onError(e: Throwable): Unit = reportError("Error in job scheduler", e)
    }
    eventLoop.start()

    // attach rate controllers of input streams to receive batch completion updates
    for {
      inputDStream <- ssc.graph.getInputStreams
      rateController <- inputDStream.rateController
    } ssc.addStreamingListener(rateController)

    listenerBus.start()
    receiverTracker = new ReceiverTracker(ssc)
    inputInfoTracker = new InputInfoTracker(ssc)

    val executorAllocClient: ExecutorAllocationClient = ssc.sparkContext.schedulerBackend match {
      case b: ExecutorAllocationClient => b.asInstanceOf[ExecutorAllocationClient]
      case _ => null
    }

    executorAllocationManager = ExecutorAllocationManager.createIfEnabled(
      executorAllocClient,
      receiverTracker,
      ssc.conf,
      ssc.graph.batchDuration.milliseconds,
      clock)
    executorAllocationManager.foreach(ssc.addStreamingListener)
    receiverTracker.start()
    jobGenerator.start()
    executorAllocationManager.foreach(_.start())
    logInfo("Started JobScheduler")
  }

在JobScheduler的start方法里面，它首先创建了EventLoop[JobSchedulerEvent]，它主要用来处理job的调度事件的，具体事件定义在processEvent里面：

 private def processEvent(event: JobSchedulerEvent) {
    try {
      event match {
        case JobStarted(job, startTime) => handleJobStart(job, startTime)
        case JobCompleted(job, completedTime) => handleJobCompletion(job, completedTime)
        case ErrorReported(m, e) => handleError(m, e)
      }
    } catch {
      case e: Throwable =>
        reportError("Error in job scheduler", e)
    }
  }

其后启动了这个eventloop，在启动之后会开启一个线程来消费eventQueue发送的事件消息，eventQueue是LinkedBlockingDeque类型的。

private val eventThread = new Thread(name) {
    setDaemon(true)

    override def run(): Unit = {
      try {
        while (!stopped.get) {
          val event = eventQueue.take()
          try {
            onReceive(event)
          } catch {
            case NonFatal(e) =>
              try {
                onError(e)
              } catch {
                case NonFatal(e) => logError("Unexpected error in " + name, e)
              }
          }
        }
      } catch {
        case ie: InterruptedException => // exit even if eventQueue is not empty
        case NonFatal(e) => logError("Unexpected error in " + name, e)
      }
    }

  }

在这个事件的接收处理启启动之后，JobScheduler启动了receiverTracker和jobGenerator，receiverTracker负责Receive分发和worker端的receive通信，并处理其发来的信息。接下来主要看jobGenerator.start的逻辑：

/** Start generation of jobs */
  def start(): Unit = synchronized {
    if (eventLoop != null) return // generator has already been started

    // Call checkpointWriter here to initialize it before eventLoop uses it to avoid a deadlock.
    // See SPARK-10125
    checkpointWriter

    eventLoop = new EventLoop[JobGeneratorEvent]("JobGenerator") {
      override protected def onReceive(event: JobGeneratorEvent): Unit = processEvent(event)

      override protected def onError(e: Throwable): Unit = {
        jobScheduler.reportError("Error in job generator", e)
      }
    }
    eventLoop.start()

    if (ssc.isCheckpointPresent) {
      restart()
    } else {
      startFirstTime()
    }
  }

在JobGenertor的start方法里面创建了EventLoop[JobGeneratorEvent]，用来处理具体的关于job的操作，具体的定义在processEvent中：

/** Processes all events */
  private def processEvent(event: JobGeneratorEvent) {
    logDebug("Got event " + event)
    event match {
      case GenerateJobs(time) => generateJobs(time)
      case ClearMetadata(time) => clearMetadata(time)
      case DoCheckpoint(time, clearCheckpointDataLater) =>
        doCheckpoint(time, clearCheckpointDataLater)
      case ClearCheckpointData(time) => clearCheckpointData(time)
    }
  }

在启动完eventloop之后，接下来会看检查点，如果第一次运行就进入startFirstTime方法中：

/** Starts the generator for the first time */
  private def startFirstTime() {
    val startTime = new Time(timer.getStartTime())
    graph.start(startTime - graph.batchDuration)
    timer.start(startTime.milliseconds)
    logInfo("Started JobGenerator at " + startTime)
  }

在startFirstTime方法里面首先会设置一个startTime，其后启动DstreamGraph，然后调用timer.start方法，timer的创建：

private val timer = new RecurringTimer(clock, ssc.graph.batchDuration.milliseconds,
    longTime => eventLoop.post(GenerateJobs(new Time(longTime))), "JobGenerator")

通过跟代码可以确定最后在线程里面运行的是triggerActionForNextInterval方法

def start(startTime: Long): Long = synchronized {
    nextTime = startTime
    thread.start()
    logInfo("Started timer for " + name + " at time " + nextTime)
    nextTime
  }

private val thread = new Thread("RecurringTimer - " + name) {
    setDaemon(true)
    override def run() { loop }
  }

private def triggerActionForNextInterval(): Unit = {
    clock.waitTillTime(nextTime)
    callback(nextTime)
    prevTime = nextTime
    nextTime += period
    logDebug("Callback for " + name + " called at time " + prevTime)
  }

  /**
   * Repeatedly call the callback every interval.
   */
  private def loop() {
    try {
      while (!stopped) {
        triggerActionForNextInterval()
      }
      triggerActionForNextInterval()
    } catch {
      case e: InterruptedException =>
    }
  }
}

在triggerActionForNextInterval方法中调用的callback方法，即timer创建的时候的eventLoop.post(GenerateJobs(new Time(longTime))方法，这里的eventloop是EventLoop[JobGeneratorEvent]类型的，所以最后会调用generateJobs方法：

/** Generate jobs and perform checkpointing for the given `time`.  */
  private def generateJobs(time: Time) {
    // Checkpoint all RDDs marked for checkpointing to ensure their lineages are
    // truncated periodically. Otherwise, we may run into stack overflows (SPARK-6847).
    ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")
    Try {
      jobScheduler.receiverTracker.allocateBlocksToBatch(time) // allocate received blocks to batch
      graph.generateJobs(time) // generate jobs using allocated block
    } match {
      case Success(jobs) =>
        val streamIdToInputInfos = jobScheduler.inputInfoTracker.getInfo(time)
        jobScheduler.submitJobSet(JobSet(time, jobs, streamIdToInputInfos))
      case Failure(e) =>
        jobScheduler.reportError("Error generating jobs for time " + time, e)
        PythonDStream.stopStreamingContextIfPythonProcessIsDead(e)
    }
    eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = false))
  }

这个方法里面先是调用jobScheduler.receiverTracker.allocateBlocksToBatch(time)方法将receive分配的block获取到这个batch中，然后在调用graph.generateJobs(time)利用上面的block来生成具体的job。接下来看jobScheduler.submitJobSet方法：

def submitJobSet(jobSet: JobSet) {
    if (jobSet.jobs.isEmpty) {
      logInfo("No jobs added for time " + jobSet.time)
    } else {
      listenerBus.post(StreamingListenerBatchSubmitted(jobSet.toBatchInfo))
      jobSets.put(jobSet.time, jobSet)
      jobSet.jobs.foreach(job => jobExecutor.execute(new JobHandler(job)))
      logInfo("Added jobs for time " + jobSet.time)
    }
  }

submitJobSet方法中会根据JobSet为每一个job新建JobHandler，放入job的线程池中，等待spark的调度处理。到此job在逻辑上已经完成。

下面是根据代码画的关于job流入线程池的时序图：

接下来看JobHandler的run方法。

def run() {
      val oldProps = ssc.sparkContext.getLocalProperties
      try {
        ssc.sparkContext.setLocalProperties(SerializationUtils.clone(ssc.savedProperties.get()))
        val formattedTime = UIUtils.formatBatchTime(
          job.time.milliseconds, ssc.graph.batchDuration.milliseconds, showYYYYMMSS = false)
        val batchUrl = s"/streaming/batch/?id=${job.time.milliseconds}"
        val batchLinkText = s"[output operation ${job.outputOpId}, batch time ${formattedTime}]"

        ssc.sc.setJobDescription(
          s"""Streaming job from <a href="$batchUrl">$batchLinkText</a>""")
        ssc.sc.setLocalProperty(BATCH_TIME_PROPERTY_KEY, job.time.milliseconds.toString)
        ssc.sc.setLocalProperty(OUTPUT_OP_ID_PROPERTY_KEY, job.outputOpId.toString)
        // Checkpoint all RDDs marked for checkpointing to ensure their lineages are
        // truncated periodically. Otherwise, we may run into stack overflows (SPARK-6847).
        ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")

        // We need to assign `eventLoop` to a temp variable. Otherwise, because
        // `JobScheduler.stop(false)` may set `eventLoop` to null when this method is running, then
        // it's possible that when `post` is called, `eventLoop` happens to null.
        var _eventLoop = eventLoop
        if (_eventLoop != null) {
          _eventLoop.post(JobStarted(job, clock.getTimeMillis()))
          // Disable checks for existing output directories in jobs launched by the streaming
          // scheduler, since we may need to write output to an existing directory during checkpoint
          // recovery; see SPARK-4835 for more details.
          SparkHadoopWriterUtils.disableOutputSpecValidation.withValue(true) {
            job.run()
          }
          _eventLoop = eventLoop
          if (_eventLoop != null) {
            _eventLoop.post(JobCompleted(job, clock.getTimeMillis()))
          }
        } else {
          // JobScheduler has been stopped.
        }
      } finally {
        ssc.sparkContext.setLocalProperties(oldProps)
      }
    }
  }

在run方法中，在调起job.run()方法运行job之后，会往evenloop发送post(JobCompleted(job, clock.getTimeMillis())这里的eventloop是EventLoop[JobSchedulerEvent]，因此具体的处理方法是handleJobCompletion：

private def handleJobCompletion(job: Job, completedTime: Long) {
    val jobSet = jobSets.get(job.time)
    jobSet.handleJobCompletion(job)
    job.setEndTime(completedTime)
    listenerBus.post(StreamingListenerOutputOperationCompleted(job.toOutputOperationInfo))
    logInfo("Finished job " + job.id + " from job set of time " + jobSet.time)
    if (jobSet.hasCompleted) {
      listenerBus.post(StreamingListenerBatchCompleted(jobSet.toBatchInfo))
    }
    job.result match {
      case Failure(e) =>
        reportError("Error running job " + job, e)
      case _ =>
        if (jobSet.hasCompleted) {
          jobSets.remove(jobSet.time)
          jobGenerator.onBatchCompletion(jobSet.time)
          logInfo("Total delay: %.3f s for time %s (execution: %.3f s)".format(
            jobSet.totalDelay / 1000.0, jobSet.time.toString,
            jobSet.processingDelay / 1000.0
          ))
        }
    }
  }

在这个方法里面根据job.result会调用（若无错误）jobGenerator.onBatchCompletion(jobSet.time)

 def onBatchCompletion(time: Time) {
    eventLoop.post(ClearMetadata(time))
  }

这个方法中eventloop发送了ClearMatadata信号，即清理元数据信号，这个信号会被EventLoop[JobGeneratorEvent]接收处理；调用claerMetadata方法

/** Clear DStream metadata for the given `time`. */
  private def clearMetadata(time: Time) {
    ssc.graph.clearMetadata(time)

    // If checkpointing is enabled, then checkpoint,
    // else mark batch to be fully processed
    if (shouldCheckpoint) {
      eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = true))
    } else {
      // If checkpointing is not enabled, then delete metadata information about
      // received blocks (block data not saved in any case). Otherwise, wait for
      // checkpointing of this batch to complete.
      val maxRememberDuration = graph.getMaxInputStreamRememberDuration()
      jobScheduler.receiverTracker.cleanupOldBlocksAndBatches(time - maxRememberDuration)
      jobScheduler.inputInfoTracker.cleanup(time - maxRememberDuration)
      markBatchFullyProcessed(time)
    }
  }

这里调用了ssc.graph.clearMetadata(time)方法：

def clearMetadata(time: Time) {
    logDebug("Clearing metadata for time " + time)
    this.synchronized {
      outputStreams.foreach(_.clearMetadata(time))
    }
    logDebug("Cleared old metadata for time " + time)
  }

上面会根据每个outputStreams来调用clearMatadata方法，这个outputstreams在DstreamGraph中定义，在调用类似foreachrdd这类触发job的算子的时候，会调用Dstream.register方法新增outputstream。最后会调用到Dstream的claerMetadata方法：

private[streaming] def clearMetadata(time: Time) {
    val unpersistData = ssc.conf.getBoolean("spark.streaming.unpersist", true)
    val oldRDDs = generatedRDDs.filter(_._1 <= (time - rememberDuration))
    logDebug("Clearing references to old RDDs: [" +
      oldRDDs.map(x => s"${x._1} -> ${x._2.id}").mkString(", ") + "]")
    generatedRDDs --= oldRDDs.keys
    if (unpersistData) {
      logDebug(s"Unpersisting old RDDs: ${oldRDDs.values.map(_.id).mkString(", ")}")
      oldRDDs.values.foreach { rdd =>
        rdd.unpersist(false)
        // Explicitly remove blocks of BlockRDD
        rdd match {
          case b: BlockRDD[_] =>
            logInfo(s"Removing blocks of RDD $b of time $time")
            b.removeBlocks()
          case _ =>
        }
      }
    }
    logDebug(s"Cleared ${oldRDDs.size} RDDs that were older than " +
      s"${time - rememberDuration}: ${oldRDDs.keys.mkString(", ")}")
    dependencies.foreach(_.clearMetadata(time))
  }

这里清理了generatedRDDs中的RDD，最后还调用了dependencies.foreach(_.clearMetadata(time))来清理数据；这个dependencies是Dstream定义的def dependencies: List[DStream[_]]，其实在Dstream的子类里面会重写，对于inputstream由于其是依赖的第一个，因此list为空，在其他Dstream中，例如MappedDStream中，其定义是list（parent）指向父类，这样依赖的关系就可以用dependencies来表示。

override def dependencies: List[DStream[_]] = List()

在项目里面用的textFileStream（）方法接收数据，其具体的实现在FileInputDstream中，在FileInputDstream中就重写了clearMetadata方法：

protected[streaming] override def clearMetadata(time: Time) {
    batchTimeToSelectedFiles.synchronized {
      val oldFiles = batchTimeToSelectedFiles.filter(_._1 < (time - rememberDuration))
      batchTimeToSelectedFiles --= oldFiles.keys
      recentlySelectedFiles --= oldFiles.values.flatten
      logInfo("Cleared " + oldFiles.size + " old files that were older than " +
        (time - rememberDuration) + ": " + oldFiles.keys.mkString(", "))
      logDebug("Cleared files are:
" +
        oldFiles.map(p => (p._1, p._2.mkString(", "))).mkString("
"))
    }
    // Delete file mod times that weren't accessed in the last round of getting new files
    fileToModTime.clearOldValues(lastNewFileFindingTime - 1)
  }

上面是FileInputDstream重写的方法，可以看到只是清理了file，但是并没有针对generatedRDDs中的RDD进行操作，因此在每一个batch结束后，由于这里的数据清理不完全，导致内存一直增加，最后OOM。这个bug在2.3.0已经修改。