本期内容：

1. 解密Spark Streaming Job架构和运行机制

2. 解密Spark Streaming 容错架构和运行机制

　　一切不能进行实时流处理的数据都是无效的数据。在流处理时代，SparkStreaming有着强大吸引力，而且发展前景广阔，加之Spark的生态系统，Streaming可以方便调用其他的诸如SQL，MLlib等强大框架，它必将一统天下。

　　Spark Streaming运行时与其说是Spark Core上的一个流式处理框架，不如说是Spark Core上的一个最复杂的应用程序。如果可以掌握Spark streaming这个复杂的应用程序，那么其他的再复杂的应用程序都不在话下了。这里选择Spark Streaming作为版本定制的切入点也是大势所趋。

本节课通过从job和容错的整体架构上来考察Spark Streaming的运行机制。

用之前已有的最简单的例子：

// Socket来源的单词计数
// YY课堂：每天20:00现场授课频道68917580
val sparkConf = new SparkConf().setMaster("local[2]").setAppName("StreamingWordCountSelfScala")
val ssc = new StreamingContext(sparkConf, Durations.seconds(5))
val lines = ssc.socketTextStream("localhost", 9999)
val words = lines.flatMap(_.split(" ")).map((_, 1)).reduceByKey(_ + _)
words.print()
ssc.start()

跟踪源码可以发现：

在初始化 StreamingContext时，创建了如下几个对象：

// StreamingContext.scala line 183
private[streaming] val scheduler = new JobScheduler(this)

而JobScheduler在初始化的时候，会初始化jobGenerator，且包含receiverTracker。

// JobScheduler.scala line 50
private val jobGenerator = new JobGenerator(this) // line 50
val clock = jobGenerator.clock
val listenerBus = new StreamingListenerBus()// These two are created only when scheduler starts.
// eventLoop not being null means the scheduler has been started and not stopped
var receiverTracker: ReceiverTracker = null // 56

再看创建DStream的部分

// StreamingContext.scala line 327
def socketTextStream(hostname: String,port: Int,storageLevel: StorageLevel = StorageLevel.MEMORY_AND_DISK_SER_2): ReceiverInputDStream[String] = withNamedScope("socket text stream") {socketStream[String](hostname, port, SocketReceiver.bytesToLines, storageLevel)
}// StreamingContext.scala line 345
def socketStream[T: ClassTag](hostname: String,port: Int,converter: (InputStream) => Iterator[T],storageLevel: StorageLevel): ReceiverInputDStream[T] = {new SocketInputDStream[T](this, hostname, port, converter, storageLevel) // line 351
}

// SocketInputDStream.scala line 33
private[streaming]
class SocketInputDStream[T: ClassTag](ssc_ : StreamingContext,host: String,port: Int,bytesToObjects: InputStream => Iterator[T],storageLevel: StorageLevel) extends ReceiverInputDStream[T](ssc_) {// 这个方法是关键def getReceiver(): Receiver[T] = {new SocketReceiver(host, port, bytesToObjects, storageLevel)}
}

再看 ssc.start

// StreamingContext.scala line 594
def start(): Unit = synchronized {state match {case INITIALIZED =>startSite.set(DStream.getCreationSite())StreamingContext.ACTIVATION_LOCK.synchronized {StreamingContext.assertNoOtherContextIsActive()try {validate()// Start the streaming scheduler in a new thread, so that thread local properties// like call sites and job groups can be reset without affecting those of the// current thread.ThreadUtils.runInNewThread("streaming-start") {sparkContext.setCallSite(startSite.get)sparkContext.clearJobGroup()sparkContext.setLocalProperty(SparkContext.SPARK_JOB_INTERRUPT_ON_CANCEL, "false")scheduler.start() // line 610}state = StreamingContextState.ACTIVE} catch {case NonFatal(e) =>logError("Error starting the context, marking it as stopped", e)scheduler.stop(false)state = StreamingContextState.STOPPEDthrow e}StreamingContext.setActiveContext(this)}shutdownHookRef = ShutdownHookManager.addShutdownHook(StreamingContext.SHUTDOWN_HOOK_PRIORITY)(stopOnShutdown)// Registering Streaming Metrics at the start of the StreamingContextassert(env.metricsSystem != null)env.metricsSystem.registerSource(streamingSource)uiTab.foreach(_.attach())logInfo("StreamingContext started")case ACTIVE =>logWarning("StreamingContext has already been started")case STOPPED =>throw new IllegalStateException("StreamingContext has already been stopped")}
}

第610行，调用了scheduler.start，scheduler就是之前初始化是产生的JobScheduler。

// JobScheduler.scala line 62
def start(): Unit = synchronized {if (eventLoop != null) return // scheduler has already been startedlogDebug("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 updatesfor {inputDStream <- ssc.graph.getInputStreamsrateController <- inputDStream.rateController} ssc.addStreamingListener(rateController)listenerBus.start(ssc.sparkContext)receiverTracker = new ReceiverTracker(ssc) // line 80inputInfoTracker = new InputInfoTracker(ssc)receiverTracker.start()jobGenerator.start()logInfo("Started JobScheduler")
}

请看80行，将receiverTracker初始化：

// ReceiverTracker.scala line 101
private[streaming]
class ReceiverTracker(ssc: StreamingContext, skipReceiverLaunch: Boolean = false) extends Logging {private val receiverInputStreams = ssc.graph.getReceiverInputStreams()private val receiverInputStreamIds = receiverInputStreams.map { _.id }private val receivedBlockTracker = new ReceivedBlockTracker(ssc.sparkContext.conf,ssc.sparkContext.hadoopConfiguration,receiverInputStreamIds,ssc.scheduler.clock,ssc.isCheckpointPresent,Option(ssc.checkpointDir))

调用receiverTracker.start和jobGenerator.star

// ReceiverTracker.scala line 148
/** Start the endpoint and receiver execution thread. */
def start(): Unit = synchronized {if (isTrackerStarted) {throw new SparkException("ReceiverTracker already started")}if (!receiverInputStreams.isEmpty) {endpoint = ssc.env.rpcEnv.setupEndpoint("ReceiverTracker", new ReceiverTrackerEndpoint(ssc.env.rpcEnv))if (!skipReceiverLaunch) launchReceivers() // line 157logInfo("ReceiverTracker started")trackerState = Started}
}

launchReceivers()

// ReceiverTracker.scala line 413
private def launchReceivers(): Unit = {val receivers = receiverInputStreams.map(nis => {val rcvr = nis.getReceiver() // 这个就是SocketInputDStream.getReceiver(),本例中是SocketReceiver ，见SocketInputDStream.scala line 34rcvr.setReceiverId(nis.id)rcvr})runDummySparkJob()logInfo("Starting " + receivers.length + " receivers")endpoint.send(StartAllReceivers(receivers)) // line 423
}

看看StartAllReceivers是如何被消费的？

// ReceiverTracker.scala line 448
// Local messages
case StartAllReceivers(receivers) =>val scheduledLocations = schedulingPolicy.scheduleReceivers(receivers, getExecutors) // 尽量负载均匀for (receiver <- receivers) {val executors = scheduledLocations(receiver.streamId)updateReceiverScheduledExecutors(receiver.streamId, executors)receiverPreferredLocations(receiver.streamId) = receiver.preferredLocationstartReceiver(receiver, executors) // 启动接收器，不再进一步深究，有兴趣的可以继续查看源码}

再回到JobScheduler.scala line 83，jobGenerator.start

// JobGenerator.scala line 79
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-10125checkpointWritereventLoop = 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()}
}

至此消息接收和Job生成器已启动。

在StreamingContext调用start方法的内部其实是会启动JobScheduler的Start方法，进行消息循环，在JobScheduler的start内部会构造JobGenerator和ReceiverTacker，并且调用JobGenerator和ReceiverTacker的start方法

　　1.JobGenerator启动后会不断的根据batchDuration生成一个个的Job

　　2.ReceiverTracker启动后首先在Spark Cluster中启动Receiver（其实是在Executor中先启动ReceiverSupervisor），在Receiver收到数据后会通过ReceiverSupervisor存储到Executor并且把数据的Metadata信息发送给Driver中的ReceiverTracker，在ReceiverTracker内部会通过ReceivedBlockTracker来管理接受到的元数据信息

　　每个BatchInterval会产生一个具体的Job，其实这里的Job不是Spark Core中所指的Job，它只是基于DStreamGraph而生成的RDD的DAG而已，从Java角度讲，相当于Runnable接口实例，此时要想运行Job需要提交给JobScheduler，在JobScheduler中通过线程池的方式找到一个单独的线程来提交Job到集群运行（其实是在线程中基于RDD的Action触发真正的作业的运行）。

　　为什么使用线程池呢？

　　 1.作业不断生成，所以为了提升效率，我们需要线程池；这和在Executor中通过线程池执行Task有异曲同工之妙；

　　2.有可能设置了Job的FAIR公平调度的方式，这个时候也需要多线程的支持。

　　第二部分：从容错架构的角度透视Spark Streaming

　　我们知道DStream与RDD的关系就是随着时间流逝不断的产生RDD，对DStream的操作就是在固定时间上操作RDD。所以从某种意义上而言，Spark Streaming的基于DStream的容错机制，实际上就是划分到每一次形成的RDD的容错机制，这也是Spark Streaming的高明之处。

　　RDD作为 分布式弹性数据集，它的弹性主要体现在：

　　1.自动的分配内存和硬盘，优先基于内存

　　2.基于lineage容错机制

　　3.task会指定次数的重试

　　4.stage失败会自动重试

　　5.checkpoint和persist 复用

　　6.数据调度弹性：DAG，TASK和资源管理无关。

　　7.数据分片的高度弹性

　　基于RDD的特性，它的容错机制主要就是两种：一是checkpoint，二是基于lineage（血统）的容错。一般而言，spark选择血统容错，因为对于大规模的数据集，做检查点的成本很高。但是有的情况下，不如说lineage链条过于复杂和冗长，这时候就需要做checkpoint。

　　考虑到RDD的依赖关系，每个stage内部都是窄依赖，此时一般基于lineage容错，方便高效。在stage之间，是宽依赖，产生了shuffle操作，这种情况下，做检查点则更好。总结来说，stage内部做lineage，stage之间做checkpoint。

后续的会有什么更深的内幕？且听下回分解。