Spark源码分析之Checkpoint的过程

概述

checkpoint 的机制保证了需要访问重复数据的应用 Spark 的DAG执行图可能很庞大，task 中计算链可能会很长，这时如果 task 中途运行出错，那么 task 的整个需要重算非常耗时，因此，有必要将计算代价较大的 RDD checkpoint 一下，当下游 RDD 计算出错时，可以直接从 checkpoint 过的 RDD 那里读取数据继续算。

我们先来看一个例子，checkpoint的使用:

import org.apache.spark.SparkContext
import org.apache.spark.SparkContext._
import org.apache.spark.SparkConf

object CheckPointTest {

   def main(args: Array[String]) {

    val sc: SparkContext = SparkContext.getOrCreate(new   SparkConf().setAppName("ck").setMaster("local[2]"))
    sc.setCheckpointDir("/Users/kinge/ck")

    val rdd: RDD[(String, Int)] = sc.textFile("").map{x=>(x,) }.reduceByKey(_+_)
    rdd.checkpoint()

    rdd.count()
    rdd.groupBy(x=>x._2).collect().foreach(println)
   }
}

checkpoint流程分析

checkpoint初始化

我们可以看到最先调用了SparkContext的setCheckpointDir 设置了一个checkpoint 目录
我们跟进这个方法看一下

/**
   * Set the directory under which RDDs are going to be checkpointed. The directory must
   * be a HDFS path if running on a cluster.
   */
  def setCheckpointDir(directory: String) {

    // If we are running on a cluster, log a warning if the directory is local.
    // Otherwise, the driver may attempt to reconstruct the checkpointed RDD from
    // its own local file system, which is incorrect because the checkpoint files
    // are actually on the executor machines.
    if (!isLocal && Utils.nonLocalPaths(directory).isEmpty) {
      logWarning("Checkpoint directory must be non-local " +
        "if Spark is running on a cluster: " + directory)
    }

   //利用hadoop的api创建了一个hdfs目录
    checkpointDir = Option(directory).map { dir =>
      val path = new Path(dir, UUID.randomUUID().toString)
      val fs = path.getFileSystem(hadoopConfiguration)
      fs.mkdirs(path)
      fs.getFileStatus(path).getPath.toString
    }
  }

这个方法挺简单的，就创建了一个目录，接下来我们看RDD核心的checkpoint 方法，跟进去

def checkpoint(): Unit = RDDCheckpointData.synchronized {
    if (context.checkpointDir.isEmpty) {
      throw new SparkException("Checkpoint directory has not been set in the SparkContext")
    } else if (checkpointData.isEmpty) {
      checkpointData = Some(new ReliableRDDCheckpointData(this))
    }
  }

这个方法没有返回值，逻辑只有一个判断，checkpointDir刚才设置过了，不为空，然后创建了一个ReliableRDDCheckpointData,我们来看ReliableRDDCheckpointData

/**
 * An implementation of checkpointing that writes the RDD data to reliable storage.
 * This allows drivers to be restarted on failure with previously computed state.
 */
private[spark] class ReliableRDDCheckpointData[T: ClassTag](@transient rdd: RDD[T])
  extends RDDCheckpointData[T](rdd) with Logging {
   。。。。。
}

这个ReliableRDDCheckpointData的父类RDDCheckpointData我们再继续看它的父类

/**
*   RDD 需要经过
*    [ Initialized  --> CheckpointingInProgress--> Checkpointed ]
*    这几个阶段才能被 checkpoint。
*/

private[spark] object CheckpointState extends Enumeration {
  type CheckpointState = Value
  val Initialized, CheckpointingInProgress, Checkpointed = Value
}

private[spark] abstract class RDDCheckpointData[T: ClassTag](@transient rdd: RDD[T])
  extends Serializable {

  import CheckpointState._

  // The checkpoint state of the associated RDD.
  protected var cpState = Initialized

  。。。。。。
}

RDD 需要经过
[ Initialized --> CheckpointingInProgress--> Checkpointed ]
这几个阶段才能被 checkpoint。
这类里面有一个枚举来标识CheckPoint的状态，第一次初始化时是Initialized。
checkpoint这个一步已经完成了，回到我们的RDD成员变量里checkpointData这个变量指向的RDDCheckpointData的实例。

Checkpoint初始化时序图:

checkpoint什么时候写入数据

我们知道一个spark job运行最终会调用SparkContext的runJob方法将任务提交给Executor去执行，我们来看runJob

def runJob[T, U: ClassTag](
      rdd: RDD[T],
      func: (TaskContext, Iterator[T]) => U,
      partitions: Seq[Int],
      resultHandler: (Int, U) => Unit): Unit = {
    if (stopped.get()) {
      throw new IllegalStateException("SparkContext has been shutdown")
    }
    val callSite = getCallSite
    val cleanedFunc = clean(func)
    logInfo("Starting job: " + callSite.shortForm)
    if (conf.getBoolean("spark.logLineage", false)) {
      logInfo("RDD's recursive dependencies:\n" + rdd.toDebugString)
    }
    dagScheduler.runJob(rdd, cleanedFunc, partitions, callSite, resultHandler, localProperties.get)
    progressBar.foreach(_.finishAll())
    rdd.doCheckpoint()
  }

最后一行代码调用了doCheckpoint，在dagScheduler将任务提交给集群运行之后，我来看这个doCheckpoint方法

private[spark] def doCheckpoint(): Unit = {
    RDDOperationScope.withScope(sc, "checkpoint", allowNesting = false, ignoreParent = true) {
      if (!doCheckpointCalled) {
        doCheckpointCalled = true
        if (checkpointData.isDefined) {
          checkpointData.get.checkpoint()
        } else {
          //遍历依赖的rdd，调用每个rdd的doCheckpoint方法
          dependencies.foreach(_.rdd.doCheckpoint())
        }
      }
    }
  }

这个是一个递归，遍历RDD依赖链条，当rdd是checkpointData不为空时，调用checkpointData的checkpoint()方法。还记得checkpointData类型是什么吗？就是RDDCheckpointData ，我们来看它的checkpoint方法，以下

final def checkpoint(): Unit = {
    // Guard against multiple threads checkpointing the same RDD by
    // atomically flipping the state of this RDDCheckpointData
    RDDCheckpointData.synchronized {
      if (cpState == Initialized) {

       //1、标记当前状态为正在checkpoint中
        cpState = CheckpointingInProgress
      } else {
        return
      }
    }

  //2 这里调用的是子类的doCheckpoint()
    val newRDD = doCheckpoint()

    // 3 标记checkpoint已完成，清空RDD依赖
    RDDCheckpointData.synchronized {
      cpRDD = Some(newRDD)
      cpState = Checkpointed
      rdd.markCheckpointed()
    }
  }

这个方法开始做checkpoint操作了，将doCheckpoint交给子类去实现checkpoint的逻辑，我们去看子类怎么实现doCheckpoint

protected override def doCheckpoint(): CheckpointRDD[T] = {

    // Create the output path for the checkpoint
    val path = new Path(cpDir)
    val fs = path.getFileSystem(rdd.context.hadoopConfiguration)
    if (!fs.mkdirs(path)) {
      throw new SparkException(s"Failed to create checkpoint path $cpDir")
    }

    //需要的配置文件（如 core-site.xml 等）broadcast 到其他 worker 节点的 blockManager。

    val broadcastedConf = rdd.context.broadcast(
      new SerializableConfiguration(rdd.context.hadoopConfiguration))

   //向集群提交一个Job去执行checkpoint操作，将RDD序列化到HDFS目录上
    rdd.context.runJob(rdd, ReliableCheckpointRDD.writeCheckpointFile[T](cpDir, broadcastedConf) _)

    // 为该 rdd 生成一个新的依赖，设置该 rdd 的 parent rdd 为
    //CheckpointRDD，该 CheckpointRDD 负责以后读取在文件系统上的
   //checkpoint 文件，生成该 rdd 的 partition。
    val newRDD = new ReliableCheckpointRDD[T](rdd.context, cpDir)
    if (newRDD.partitions.length != rdd.partitions.length) {
      throw new SparkException(
        s"Checkpoint RDD $newRDD(${newRDD.partitions.length}) has different " +
          s"number of partitions from original RDD $rdd(${rdd.partitions.length})")
    }

    // 是否清除checkpoint文件如果超出引用的资源范围
    if (rdd.conf.getBoolean("spark.cleaner.referenceTracking.cleanCheckpoints", false)) {
      rdd.context.cleaner.foreach { cleaner =>
        cleaner.registerRDDCheckpointDataForCleanup(newRDD, rdd.id)
      }
    }

    logInfo(s"Done checkpointing RDD ${rdd.id} to $cpDir, new parent is RDD ${newRDD.id}")

//  将新产生的RDD返回给父类
    newRDD
  }

上面的代码最终会返回新的CheckpointRDD ，父类将它赋值给成员变量cpRDD，最终标记当前状态为Checkpointed并清空当前RDD的依赖链。到此Checkpoint的数据就被序列化到HDFS上了。

Checkpoint 写数据时序图

checkpoint什么时候读取数据

我们知道Task是spark运行任务的最小单元，当Task执行失败的时候spark会重新计算，这里Task进行计算的地方就是读取checkpoint的入口。我们可以看一下ShuffleMapTask 里的计算方法runTask,如下

override def runTask(context: TaskContext): MapStatus = {

     。。。。。。。

    try {
      val manager = SparkEnv.get.shuffleManager
      writer = manager.getWriter[Any, Any](dep.shuffleHandle, partitionId, context)

    //调用rdd.iterator，迭代每个partition里的数据，计算并写入磁盘
      writer.write(rdd.iterator(partition, context).asInstanceOf[Iterator[_ <: Product2[Any, Any]]])

      writer.stop(success = true).get
    } catch {
      case e: Exception =>
        try {
          if (writer != null) {
            writer.stop(success = false)
          }
        } catch {
          case e: Exception =>
            log.debug("Could not stop writer", e)
        }
        throw e
    }
  }

这是spark真正调用计算方法的逻辑runTask调用 rdd.iterator() 去计算该 rdd 的 partition 的，我们来看RDD的iterator()

final def iterator(split: Partition, context: TaskContext): Iterator[T] = {
    if (storageLevel != StorageLevel.NONE) {
      SparkEnv.get.cacheManager.getOrCompute(this, split, context, storageLevel)
    } else {
      computeOrReadCheckpoint(split, context)
    }
  }

这里会继续调用computeOrReadCheckpoint,我们看该方法

**
   * Compute an RDD partition or read it from a checkpoint if the RDD is checkpointing.
   */
  private[spark] def computeOrReadCheckpoint(split: Partition, context: TaskContext): Iterator[T] =
  {
    if (isCheckpointedAndMaterialized) {
      firstParent[T].iterator(split, context)
    } else {
      compute(split, context)
    }
  }

当调用rdd.iterator()去计算该 rdd 的 partition 的时候，会调用 computeOrReadCheckpoint(split: Partition)去查看该 rdd 是否被 checkpoint 过了，如果是，就调用该 rdd 的 parent rdd 的 iterator() 也就是 CheckpointRDD.iterator()，否则直接调用该RDD的compute, 那么我们就跟进CheckpointRDD的compute

/**
   * Read the content of the checkpoint file associated with the given partition.
   */
  override def compute(split: Partition, context: TaskContext): Iterator[T] = {
    val file = new Path(checkpointPath, ReliableCheckpointRDD.checkpointFileName(split.index))
    ReliableCheckpointRDD.readCheckpointFile(file, broadcastedConf, context)
  }

这里就两行代码，意思是从Path上读取我们的CheckPoint数据，看一下readCheckpointFile

/**
   * Read the content of the specified checkpoint file.
   */
  def readCheckpointFile[T](
      path: Path,
      broadcastedConf: Broadcast[SerializableConfiguration],
      context: TaskContext): Iterator[T] = {
    val env = SparkEnv.get

  // 用hadoop API 读取HDFS上的数据
    val fs = path.getFileSystem(broadcastedConf.value.value)
    val bufferSize = env.conf.getInt("spark.buffer.size", )
    val fileInputStream = fs.open(path, bufferSize)
    val serializer = env.serializer.newInstance()
    val deserializeStream = serializer.deserializeStream(fileInputStream)

    // Register an on-task-completion callback to close the input stream.
    context.addTaskCompletionListener(context => deserializeStream.close())

    //反序列化数据后转换为一个Iterator
    deserializeStream.asIterator.asInstanceOf[Iterator[T]]

CheckpointRDD 负责读取文件系统上的文件，生成该 rdd 的 partition。这就解释了为什么要为调用了checkpoint的RDD 添加一个 parent CheckpointRDD的原因。
到此，整个checkpoint的流程就结束了。

Checkpoint 读取数据时序图