第12课：Spark Streaming源码解读之executor容错安全性

本期内容：

1.Executor的WAL

2.消息重放

3.其它

 

StorageLevel.scala

Memory不够的时候才考虑disk

classStorageLevelprivate(
    private var _useDisk: Boolean,
    private var _useMemory: Boolean,
    private var _useOffHeap: Boolean,
    private var _deserialized: Boolean,
    private var _replication: Int =1)
  extends Externalizable {

ReceiverSupervisorImpl.scala

/** Store the bytes of received data as a data block into Spark's memory. */
def pushBytes(
    bytes: ByteBuffer,
    metadataOption: Option[Any],
    blockIdOption: Option[StreamBlockId]
  ) {
  pushAndReportBlock(ByteBufferBlock(bytes), metadataOption, blockIdOption)
}

/** Store block and report it to driver */
def pushAndReportBlock(
    receivedBlock: ReceivedBlock,
    metadataOption: Option[Any],
    blockIdOption: Option[StreamBlockId]
  ) {
  val blockId = blockIdOption.getOrElse(nextBlockId)
  val time = System.currentTimeMillis
  val blockStoreResult =receivedBlockHandler.storeBlock(blockId, receivedBlock)
  logDebug(s"Pushed block $blockId in${(System.currentTimeMillis- time)} ms")
  val numRecords = blockStoreResult.numRecords
  val blockInfo =ReceivedBlockInfo(streamId, numRecords, metadataOption, blockStoreResult)
  trackerEndpoint.askWithRetry[Boolean](AddBlock(blockInfo))
  logDebug(s"Reported block $blockId")
}

receivedBlockHandler

private valreceivedBlockHandler: ReceivedBlockHandler = {
  if (WriteAheadLogUtils.enableReceiverLog(env.conf)) {
    if (checkpointDirOption.isEmpty) {
      throw new SparkException(
        "Cannot enable receiver write-ahead log without checkpoint directory set. "+
          "Please use streamingContext.checkpoint() to set the checkpoint directory. "+
          "See documentation for more details.")
    }
    new WriteAheadLogBasedBlockHandler(env.blockManager, receiver.streamId,
      receiver.storageLevel, env.conf, hadoopConf, checkpointDirOption.get)
  } else {
    new BlockManagerBasedBlockHandler(env.blockManager, receiver.storageLevel)
  }
}

def enableReceiverLog(conf: SparkConf): Boolean = {
  conf.getBoolean(RECEIVER_WAL_ENABLE_CONF_KEY,false)
}

ReceivedBlockHandler

def storeBlock(blockId: StreamBlockId, block: ReceivedBlock): ReceivedBlockStoreResult = {

  var numRecords = None: Option[Long]

  val putResult:Seq[(BlockId, BlockStatus)] = blockmatch {
    case ArrayBufferBlock(arrayBuffer) =>
      numRecords = Some(arrayBuffer.size.toLong)
      blockManager.putIterator(blockId, arrayBuffer.iterator, storageLevel,
        tellMaster = true)
    case IteratorBlock(iterator) =>
      val countIterator =new CountingIterator(iterator)
      val putResult = blockManager.putIterator(blockId, countIterator, storageLevel,
        tellMaster = true)
      numRecords = countIterator.count
      putResult
    case ByteBufferBlock(byteBuffer) =>
      blockManager.putBytes(blockId, byteBuffer, storageLevel, tellMaster =true)
    case o =>
      throw new SparkException(
        s"Could not store $blockId to block manager, unexpected block type${o.getClass.getName}")
  }
  if (!putResult.map { _._1 }.contains(blockId)) {
    throw new SparkException(
      s"Could not store $blockId to block manager with storage level$storageLevel")
  }
  BlockManagerBasedStoreResult(blockId, numRecords)
}

def putIterator(
    blockId: BlockId,
    values: Iterator[Any],
    level: StorageLevel,
    tellMaster: Boolean = true,
    effectiveStorageLevel: Option[StorageLevel] = None): Seq[(BlockId, BlockStatus)] = {
  require(values != null,"Values is null")
  doPut(blockId, IteratorValues(values), level, tellMaster, effectiveStorageLevel)
}

FileBasedWriteAheadLog

write

/**
 * Write a byte buffer to the log file. This method synchronously writes the data in the
 * ByteBuffer to HDFS. When this method returns, the data is guaranteed to have been flushed
 * to HDFS, and will be available for readers to read.
 */
def write(byteBuffer: ByteBuffer, time: Long): FileBasedWriteAheadLogSegment = synchronized {
  var fileSegment:FileBasedWriteAheadLogSegment = null
  var failures = 0
  var lastException:Exception = null
  var succeeded = false
  while (!succeeded && failures < maxFailures) {
    try {
      fileSegment = getLogWriter(time).write(byteBuffer)
      if (closeFileAfterWrite) {
        resetWriter()
      }
      succeeded = true
    } catch{
      case ex:Exception =>
        lastException = ex
        logWarning("Failed to write to write ahead log")
        resetWriter()
        failures += 1
    }
  }
  if (fileSegment ==null) {
    logError(s"Failed to write to write ahead log after$failures failures")
    throw lastException
  }
  fileSegment
}

/** Get the current log writer while taking care of rotation */
private def getLogWriter(currentTime: Long): FileBasedWriteAheadLogWriter = synchronized {
  if (currentLogWriter== null|| currentTime > currentLogWriterStopTime) {
    resetWriter()
    currentLogPath.foreach {
      pastLogs +=LogInfo(currentLogWriterStartTime,currentLogWriterStopTime, _)
    }
    currentLogWriterStartTime = currentTime
    currentLogWriterStopTime = currentTime + (rollingIntervalSecs *1000)
    val newLogPath =new Path(logDirectory,
      timeToLogFile(currentLogWriterStartTime,currentLogWriterStopTime))
    currentLogPath =Some(newLogPath.toString)
    currentLogWriter =new FileBasedWriteAheadLogWriter(currentLogPath.get, hadoopConf)
  }
  currentLogWriter
}

readAll

def readAll(): JIterator[ByteBuffer] = synchronized {
  val logFilesToRead =pastLogs.map{ _.path} ++currentLogPath
  logInfo("Reading from the logs:\n"+ logFilesToRead.mkString("\n"))
  def readFile(file:String): Iterator[ByteBuffer] = {
    logDebug(s"Creating log reader with $file")
    val reader =new FileBasedWriteAheadLogReader(file, hadoopConf)
    CompletionIterator[ByteBuffer, Iterator[ByteBuffer]](reader, reader.close _)
  }
  if (!closeFileAfterWrite) {
    logFilesToRead.iterator.map(readFile).flatten.asJava
  } else {
    // For performance gains, it makes sense to parallelize the recovery if
    // closeFileAfterWrite = true
    seqToParIterator(threadpool, logFilesToRead, readFile).asJava
  }
}

FileBasedWriteAheadLogReader

private[streaming]class FileBasedWriteAheadLogReader(path:String, conf: Configuration)
  extends Iterator[ByteBuffer]with Closeablewith Logging {

private val instream= HdfsUtils.getInputStream(path, conf)
private var closed= (instream== null)// the file may be deleted as we're opening the stream
private var nextItem: Option[ByteBuffer] =None

 

override defhasNext: Boolean = synchronized {
  if (closed) {
    return false
  }

  if (nextItem.isDefined) {// handle the case where hasNext is called without calling next
    true
  } else{
    try {
      val length =instream.readInt()
      val buffer =new Array[Byte](length)
      instream.readFully(buffer)
      nextItem =Some(ByteBuffer.wrap(buffer))
      logTrace("Read next item " + nextItem.get)
      true
    } catch{
      case e: EOFException =>
        logDebug("Error reading next item, EOF reached", e)
        close()
        false
      case e: IOException =>
        logWarning("Error while trying to read data. If the file was deleted, "+
          "this should be okay.", e)
        close()
        if (HdfsUtils.checkFileExists(path, conf)) {
          // If file exists, this could be a legitimate error
          throw e
        } else {
          // File was deleted. This can occur when the daemon cleanup thread takes time to
          // delete the file during recovery.
          false
        }

      case e:Exception =>
        logWarning("Error while trying to read data from HDFS.", e)
        close()
        throw e
    }
  }
}

getInputStream

def getInputStream(path: String, conf: Configuration): FSDataInputStream = {
  val dfsPath =new Path(path)
  val dfs =getFileSystemForPath(dfsPath, conf)
  if (dfs.isFile(dfsPath)) {
    try {
      dfs.open(dfsPath)
    } catch {
      case e: IOException =>
        // If we are really unlucky, the file may be deleted as we're opening the stream.
        // This can happen as clean up is performed by daemon threads that may be left over from
        // previous runs.
        if (!dfs.isFile(dfsPath))null else throw e
    }
  } else {
    null
  }
}

DirectKafkaInputDStream

@tailrec
protected final def latestLeaderOffsets(retries: Int): Map[TopicAndPartition, LeaderOffset] = {
  val o = kc.getLatestLeaderOffsets(currentOffsets.keySet)
  // Either.fold would confuse @tailrec, do it manually
  if (o.isLeft) {
    val err = o.left.get.toString
    if (retries <=0) {
      throw new SparkException(err)
    } else {
      log.error(err)
      Thread.sleep(kc.config.refreshLeaderBackoffMs)
      latestLeaderOffsets(retries - 1)
    }
  } else {
    o.right.get
  }
}

KafkaRDD

private[kafka]
class KafkaRDD[
  K: ClassTag,
  V: ClassTag,
  U <: Decoder[_]: ClassTag,
  T <: Decoder[_]: ClassTag,
  R: ClassTag] private[spark] (
    sc: SparkContext,
    kafkaParams: Map[String, String],
    val offsetRanges: Array[OffsetRange],
    leaders: Map[TopicAndPartition, (String, Int)],
    messageHandler: MessageAndMetadata[K, V] => R
  ) extends RDD[R](sc, Nil)with Loggingwith HasOffsetRanges {
  override def getPartitions: Array[Partition] = {
    offsetRanges.zipWithIndex.map { case (o, i) =>
        val (host, port) = leaders(TopicAndPartition(o.topic, o.partition))
        new KafkaRDDPartition(i, o.topic, o.partition, o.fromOffset, o.untilOffset, host, port)
    }.toArray
  }

getPreferredLocations

override defgetPreferredLocations(thePart: Partition): Seq[String] = {
  val part = thePart.asInstanceOf[KafkaRDDPartition]
  // TODO is additional hostname resolution necessary here
  Seq(part.host)
}

connectLeader

// The idea is to use the provided preferred host, except on task retry atttempts,
// to minimize number of kafka metadata requests
private def connectLeader: SimpleConsumer = {
  if (context.attemptNumber >0) {
    kc.connectLeader(part.topic, part.partition).fold(
      errs => throw new SparkException(
        s"Couldn't connect to leader for topic${part.topic}${part.partition}: "+
          errs.mkString("\n")),
      consumer => consumer
    )
  } else {
    kc.connect(part.host, part.port)
  }
}