Spark Checkpoint读操作代码分析

Spark 2015-12-23 3935 0评论 下载为PDF    为什么不允许复制

　　《Spark RDD缓存代码分析》
　　《Spark Task序列化代码分析》
　　《Spark分区器HashPartitioner和RangePartitioner代码详解》
　　《Spark Checkpoint读操作代码分析》
　　《Spark Checkpoint写操作代码分析》

　　上次介绍了RDD的Checkpint写过程（《Spark Checkpoint写操作代码分析》），本文将介绍RDD如何读取已经Checkpint的数据。在RDD Checkpint完之后，Checkpint的信息（比如数据存放的目录）都由RDDCheckpointData去管理，所以当下次计算依赖了这个RDD的时候，首先是根据依赖关系判断出当前这个RDD是否被Checkpint了，主要是通过RDD的dependencies决定：

final def dependencies: Seq[Dependency[_]] = {

  checkpointRDD.map(r => List(new OneToOneDependency(r))).getOrElse {

    if (dependencies_ == null) {

      dependencies_ = getDependencies

    }

    dependencies_

  }

}

　　如果RDD被Checkpint了，那么checkpointRDD为Some(CheckpointRDD[T])了，所以依赖的RDD变成了CheckpointRDD。在计算数据的过程中会调用RDD的iterator方法：

final def iterator(split: Partition, context: TaskContext): Iterator[T] = {

  if (storageLevel != StorageLevel.NONE) {

    <span class="wp_keywordlink_affiliate"><a data-original-title="View all posts in Spark" href="https://www.iteblog.com/archives/tag/spark/" title="" target="_blank">Spark</a></span>Env.get.cacheManager.getOrCompute(this, split, context, storageLevel)

  } else {

    computeOrReadCheckpoint(split, context)

  }

}

 

private[spark] def computeOrReadCheckpoint(split: Partition, context: TaskContext): Iterator[T] =

{

   if (isCheckpointed) firstParent[T].iterator(split, context) else compute(split, context)

}

　　计算的过程中首先会判断RDD是否被Checkpint了，而RDD Checkpint写之后这个条件肯定是true的。而firstParent已经变成了CheckpointRDD，所以会调用CheckpointRDD的iterator方法， 该方法最终会调用ReliableCheckpointRDD的compute方法：

override def compute(split: Partition, context: TaskContext): Iterator[T] = {

  val file = new Path(checkpointPath, ReliableCheckpointRDD.checkpointFileName(split.index))

  ReliableCheckpointRDD.readCheckpointFile(file, broadcastedConf, context)

}

　　在compute方法中会通过ReliableCheckpointRDD的readCheckpointFile方法来从file路径里面读出已经Checkpint的数据，readCheckpointFile的实现如下：

def readCheckpointFile[T](

    path: Path,

    broadcastedConf: Broadcast[SerializableConfiguration],

    context: TaskContext): Iterator[T] = {

  val env = <span class="wp_keywordlink_affiliate"><a data-original-title="View all posts in Spark" href="https://www.iteblog.com/archives/tag/spark/" title="" target="_blank">Spark</a></span>Env.get

  val fs = path.getFileSystem(broadcastedConf.value.value)

  val bufferSize = env.conf.getInt("spark.buffer.size", 65536)

  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())

 

  deserializeStream.asIterator.asInstanceOf[Iterator[T]]

}

最后数据就回被全部读取出来，整个Checkpint读过程完成了。