16.Spark Streaming源码解读之数据清理机制解析

原创文章，转载请注明:转载自 听风居士博客(http://www.cnblogs.com/zhouyf/)

本期内容：

一、Spark Streaming 数据清理总览

二、Spark Streaming 数据清理过程详解

三、Spark Streaming 数据清理的触发机制


    Spark Streaming不像普通Spark 的应用程序，普通Spark程序运行完成后,中间数据会随着SparkContext的关闭而被销毁,而Spark Streaming一直在运行，不断计算，每一秒中在不断运行都会产生大量的中间数据，所以需要对对象及元数据需要定期清理。每个batch duration运行时不断触发job后需要清理rdd和元数据。下面我们就结合源码详细解析一下Spark Streaming程序的数据清理机制。


一、数据清理总览

    Spark Streaming 运行过程中,随着时间不断产生Job,当job运行结束后,需要清理相应的数据(RDD,元数据信息,Checkpoint数据),Job由JobGenerator定时产生,数据的清理也是有JobGenerator负责。

    JobGenerator负责数据清理控制的代码位于一个消息循环体eventLoop中：

  1.     eventLoop = new EventLoop[JobGeneratorEvent]("JobGenerator") {
  2.       override protected def onReceive(event: JobGeneratorEvent): Unit = processEvent(event)
  4.       override protected def onError(e: Throwable): Unit = {
  5.         jobScheduler.reportError("Error in job generator", e)
  6.       }
  7.     }
  8.     eventLoop.start()

其中的核心逻辑位于processEvent(event)函数中:

  1.   /** Processes all events */
  2.   private def processEvent(event: JobGeneratorEvent) {
  3.     logDebug("Got event " + event)
  4.     event match {
  5.       case GenerateJobs(time) => generateJobs(time)
  6.       case ClearMetadata(time) => clearMetadata(time)
  7.       case DoCheckpoint(time, clearCheckpointDataLater) =>
  8.         doCheckpoint(time, clearCheckpointDataLater)
  9.       case ClearCheckpointData(time) => clearCheckpointData(time)
  10.     }
  11.   }

可以看到当JobGenerator收到ClearMetadata(time) 和 ClearCheckpointData(time)是会进行相应的数据清理,其中 clearMetadata(time)会清理RDD数据和一些元数据信息, ClearCheckpointData(time)会清理Checkpoint数据。

二、数据清理过程详解
    2.1 ClearMetaData 过程详解
首先看一下clearMetaData函数的处理逻辑：
  1.   /** Clear DStream metadata for the given `time`. */
  2.   private def clearMetadata(time: Time) {
  3.     ssc.graph.clearMetadata(time)
  5.     // If checkpointing is enabled, then checkpoint,
  6.     // else mark batch to be fully processed
  7.     if (shouldCheckpoint) {
  8.       eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = true))
  9.     } else {
  10.       // If checkpointing is not enabled, then delete metadata information about
  11.       // received blocks (block data not saved in any case). Otherwise, wait for
  12.       // checkpointing of this batch to complete.
  13.       val maxRememberDuration = graph.getMaxInputStreamRememberDuration()
  14.       jobScheduler.receiverTracker.cleanupOldBlocksAndBatches(time - maxRememberDuration)
  15.       jobScheduler.inputInfoTracker.cleanup(time - maxRememberDuration)
  16.       markBatchFullyProcessed(time)
  17.     }
  18.   }

首先调用了DStreamGraph的clearMetadata方法:

  1.   def clearMetadata(time: Time) {
  2.     logDebug("Clearing metadata for time " + time)
  3.     this.synchronized {
  4.       outputStreams.foreach(_.clearMetadata(time))
  5.     }
  6.     logDebug("Cleared old metadata for time " + time)
  7.   }

这里调用了所有OutputDStream (关于DStream 的分类请参考http://blog.csdn.net/zhouzx2010/article/details/51460790)的clearMetadata方法

  1.   private[streaming] def clearMetadata(time: Time) {
  2.     val unpersistData = ssc.conf.getBoolean("spark.streaming.unpersist", true)
  3.      //获取需要清理的RDD
  4.     val oldRDDs = generatedRDDs.filter(_._1 <= (time - rememberDuration))
  5.     logDebug("Clearing references to old RDDs: [" +
  6.       oldRDDs.map(x => s"${x._1} -> ${x._2.id}").mkString(", ") + "]")

  8.     //将要清除的RDD从generatedRDDs 中清除 
  9.     generatedRDDs --= oldRDDs.keys
  10.     if (unpersistData) {
  11.       logDebug(s"Unpersisting old RDDs: ${oldRDDs.values.map(_.id).mkString(", ")}")
  12.       oldRDDs.values.foreach { rdd =>
  13.       //将RDD 从persistence列表中移除
  14.         rdd.unpersist(false)
  15.         // Explicitly remove blocks of BlockRDD
  16.         rdd match {
  17.           case b: BlockRDD[_] =>
  18.             logInfo(s"Removing blocks of RDD $b of time $time")
  19.             //移除RDD的block 数据
  20.             b.removeBlocks()
  21.           case _ =>
  22.         }
  23.       }
  24.     }
  25.     logDebug(s"Cleared ${oldRDDs.size} RDDs that were older than " +
  26.       s"${time - rememberDuration}: ${oldRDDs.keys.mkString(", ")}")
  27.     //清除依赖的DStream
  28.     dependencies.foreach(_.clearMetadata(time))
  29.   }

关键的清理逻辑在代码中做了详细注释,首先清理DStream对应的RDD的元数据信息,然后清理RDD的数据,最后对DStream所依赖的DStream进行清理。

回到JobGenerator的clearMetadata函数：

  1.     if (shouldCheckpoint) {
  2.       eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = true))
  3.     } else {
  4.       // If checkpointing is not enabled, then delete metadata information about
  5.       // received blocks (block data not saved in any case). Otherwise, wait for
  6.       // checkpointing of this batch to complete.
  7.       val maxRememberDuration = graph.getMaxInputStreamRememberDuration()
  8.       jobScheduler.receiverTracker.cleanupOldBlocksAndBatches(time - maxRememberDuration)
  9.       jobScheduler.inputInfoTracker.cleanup(time - maxRememberDuration)
  10.       markBatchFullyProcessed(time)
  11.     }

调用了ReceiverTracker的 cleanupOldBlocksAndBatches方法，最后调用了clearupOldBatches方法：

  1.   def cleanupOldBatches(cleanupThreshTime: Time, waitForCompletion: Boolean): Unit = synchronized {
  2.     require(cleanupThreshTime.milliseconds < clock.getTimeMillis())
  3.     val timesToCleanup = timeToAllocatedBlocks.keys.filter { _ < cleanupThreshTime }.toSeq
  4.     logInfo(s"Deleting batches: ${timesToCleanup.mkString(" ")}")
  5.     if (writeToLog(BatchCleanupEvent(timesToCleanup))) {
  6.       //将要删除的Batch数据清除
  7.       timeToAllocatedBlocks --= timesToCleanup
  8.       //清理WAL日志
  9.       writeAheadLogOption.foreach(_.clean(cleanupThreshTime.milliseconds, waitForCompletion))
  10.     } else {
  11.       logWarning("Failed to acknowledge batch clean up in the Write Ahead Log.")
  12.     }
  13.   }

可以看到ReceiverTracker的clearupOldBatches方法清理了Receiver数据,也就是Batch数据和WAL日志数据。
最后对InputInfoTracker信息进行清理：
  1.   def cleanup(batchThreshTime: Time): Unit = synchronized {
  2.     val timesToCleanup = batchTimeToInputInfos.keys.filter(_ < batchThreshTime)
  3.     logInfo(s"remove old batch metadata: ${timesToCleanup.mkString(" ")}")
  4.     batchTimeToInputInfos --= timesToCleanup
  5.   }
这简单的清除了batchTimeToInputInfos 的输入信息。

2.2 ClearCheckPoint 过程详解
看一下clearCheckpointData的处理逻辑:
  1.   /** Clear DStream checkpoint data for the given `time`. */
  2.   private def clearCheckpointData(time: Time) {
  3.     ssc.graph.clearCheckpointData(time)
  5.     // All the checkpoint information about which batches have been processed, etc have
  6.     // been saved to checkpoints, so its safe to delete block metadata and data WAL files
  7.     val maxRememberDuration = graph.getMaxInputStreamRememberDuration()
  8.     jobScheduler.receiverTracker.cleanupOldBlocksAndBatches(time - maxRememberDuration)
  9.     jobScheduler.inputInfoTracker.cleanup(time - maxRememberDuration)
  10.     markBatchFullyProcessed(time)
  11.   }

后面的ReceiverTraker和InputInforTracker的清理逻辑和ClearMetaData的相同,这分析DStreamGraph的clearCheckpointData方法:

  1.   def clearCheckpointData(time: Time) {
  2.     logInfo("Clearing checkpoint data for time " + time)
  3.     this.synchronized {
  4.       outputStreams.foreach(_.clearCheckpointData(time))
  5.     }
  6.     logInfo("Cleared checkpoint data for time " + time)
  7.   }

同样的调用了DStreamGraph中所有OutputDStream的clearCheckPiontData 方法:

  1.   private[streaming] def clearCheckpointData(time: Time) {
  2.     logDebug("Clearing checkpoint data")
  3.     checkpointData.cleanup(time)
  4.     dependencies.foreach(_.clearCheckpointData(time))
  5.     logDebug("Cleared checkpoint data")
  6.   }

这里的核心逻辑在checkpointData.cleanup(time)方法,这里的CheckpointData 是 DStreamCheckpointData对象, DStreamCheckpointData的clearup方法如下:
  1. def cleanup(time: Time) {
  2.     // 获取需要清理的Checkpoint 文件 时间
  3.     timeToOldestCheckpointFileTime.remove(time) match {
  4.       case Some(lastCheckpointFileTime) =>
  5.         //获取需要删除的文件
  6.         val filesToDelete = timeToCheckpointFile.filter(_._1 < lastCheckpointFileTime)
  7.         logDebug("Files to delete:\n" + filesToDelete.mkString(","))
  8.         filesToDelete.foreach {
  9.           case (time, file) =>
  10.             try {
  11.               val path = new Path(file)
  12.               if (fileSystem == null) {
  13.                 fileSystem = path.getFileSystem(dstream.ssc.sparkContext.hadoopConfiguration)
  14.               }
  15.           //  删除文件     
  16.               fileSystem.delete(path, true)
  17.               timeToCheckpointFile -= time
  18.               logInfo("Deleted checkpoint file '" + file + "' for time " + time)
  19.             } catch {
  20.               case e: Exception =>
  21.                 logWarning("Error deleting old checkpoint file '" + file + "' for time " + time, e)
  22.                 fileSystem = null
  23.             }
  24.         }
  25.       case None =>
  26.         logDebug("Nothing to delete")
  27.     }
  28.   }

可以看到checkpoint的清理,就是删除了指定时间以前的checkpoint文件。

三、数据清理的触发
    3.1 ClearMetaData 过程的触发
JobGenerator 生成job后,交给JobHandler执行, JobHandler的run方法中,会在job执行完后给JobScheduler 发送JobCompleted消息:
  1.           _eventLoop = eventLoop
  2.           if (_eventLoop != null) {
  3.             _eventLoop.post(JobCompleted(job, clock.getTimeMillis()))
  5.           }
JobScheduler 收到JobCompleted 消息调用 handleJobCompletion 方法,源码如下:

  1.   private def processEvent(event: JobSchedulerEvent) {
  2.     try {
  3.       event match {
  4.         case JobStarted(job, startTime) => handleJobStart(job, startTime)
  5.         case JobCompleted(job, completedTime) => handleJobCompletion(job, completedTime)
  6.         case ErrorReported(m, e) => handleError(m, e)
  7.       }
  8.     } catch {
  9.       case e: Throwable =>
  10.         reportError("Error in job scheduler", e)
  11.     }
  12.   }

在 JobScheduler 的handleJobCompletion方法中会调用JobGenerator的onBatchCompletion方法,我们看一下JobGenerator的 onBatchCompletion 方法的源码:

  1.   def onBatchCompletion(time: Time) {
  2.     eventLoop.post(ClearMetadata(time))
  3.   }

可以看到JobGenerator的onBatchCompletion方法给自己发送了ClearMetadata消息从而触发了ClearMetaData操作。

3.2 ClearCheckPoint 过程的触发
    清理CheckPoint数据发生在CheckPoint完成之后,我们先看一下CheckPointHandler的run方法:

  1.           // All done, print success
  2.           val finishTime = System.currentTimeMillis()
  3.           logInfo("Checkpoint for time " + checkpointTime + " saved to file '" + checkpointFile +
  4.             "', took " + bytes.length + " bytes and " + (finishTime - startTime) + " ms")
  5.           //调用JobGenerator的方法进行checkpoint数据清理
  6.           jobGenerator.onCheckpointCompletion(checkpointTime, clearCheckpointDataLater)
  7.           return

可以看到在checkpoint完成后,会调用JobGenerator的onCheckpointCompletion方法进行checkpoint数据清理,我查看JobGenerator的onCheckpointCompletion方法源码:

  1.   def onCheckpointCompletion(time: Time, clearCheckpointDataLater: Boolean) {
  2.     if (clearCheckpointDataLater) {
  3.       eventLoop.post(ClearCheckpointData(time))
  4.     }
  5.   }

可以看到JobGenerator的onCheckpointCompletion方法中首先对传进来的 clearCheckpointDataLater 参数进行判断,如果该参数为true,就会给JobGenerator的eventLoop循环体发送ClearCheckpointData消息,从而触发clearCheckpointData 方法的调用,进行Checkpoint数据的清理。
什么时候该参数会true呢？
我们回到JobGenerator的 ClearMetadata 方法：

  1.   private def clearMetadata(time: Time) {
  2.     ssc.graph.clearMetadata(time)

  5.     if (shouldCheckpoint) {
  6.       //发送DoCheckpoint消息,并进行相应的Checkpoint数据清理
  7.       eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = true))
  8.     } else {

  10.       val maxRememberDuration = graph.getMaxInputStreamRememberDuration()
  11.       jobScheduler.receiverTracker.cleanupOldBlocksAndBatches(time - maxRememberDuration)
  12.       jobScheduler.inputInfoTracker.cleanup(time - maxRememberDuration)
  13.       markBatchFullyProcessed(time)
  14.     }
  15.   }

可以看到在clearMetadata方法中,发送了DoCheckpoint消息,其中参数 clearCheckpointDataLater 为ture。Generator的eventLoop收到该消息后调用 doCheckpoint 方法：

  1.   private def doCheckpoint(time: Time, clearCheckpointDataLater: Boolean) {
  2.     if (shouldCheckpoint && (time - graph.zeroTime).isMultipleOf(ssc.checkpointDuration)) {
  3.       logInfo("Checkpointing graph for time " + time)
  4.       ssc.graph.updateCheckpointData(time)
  5.       checkpointWriter.write(new Checkpoint(ssc, time), clearCheckpointDataLater)
  6.     }
  7.   }

这里关键一步：调用了CheckpointWriter的write方法,注意此时参数 clearCheckpointDataLater 为true。我们进入该方法：

  1.   def write(checkpoint: Checkpoint, clearCheckpointDataLater: Boolean) {
  2.     try {
  3.       val bytes = Checkpoint.serialize(checkpoint, conf)
  4.      //将参数clearCheckpointDataLater传入CheckpoitWriteHandler
  5.       executor.execute(new CheckpointWriteHandler(
  6.         checkpoint.checkpointTime, bytes, clearCheckpointDataLater))
  7.       logInfo("Submitted checkpoint of time " + checkpoint.checkpointTime + " writer queue")
  8.     } catch {
  9.       case rej: RejectedExecutionException =>
  10.         logError("Could not submit checkpoint task to the thread pool executor", rej)
  11.     }
  12.   }

可以看到此时参数 clearCheckpointDataLater 传入CheckpointWriteHandler 。这样Checkpoint完成之后就会发送ClearCheckpointData消息给JobGenerator进行Checkpoint数据的清理。

原创文章，转载请注明:转载自 听风居士博客(http://www.cnblogs.com/zhouyf/)










 



posted on 2016-07-01 14:43  听风居士  阅读(744)  评论(0)    收藏  举报

导航