Spark Streaming Backpressure反压机制

Spark Streaming Backpressure分析

 

为什么引入Backpressure

默认情况下，Spark Streaming通过Receiver以生产者生产数据的速率接收数据，计算过程中会出现batch processing time > batch interval的情况，其中batch processing time 为实际计算一个批次花费时间， batch interval为Streaming应用设置的批处理间隔。这意味着Spark Streaming的数据接收速率高于Spark从队列中移除数据的速率，也就是数据处理能力低，在设置间隔内不能完全处理当前接收速率接收的数据。如果这种情况持续过长的时间，会造成数据在内存中堆积，导致Receiver所在Executor内存溢出等问题（如果设置StorageLevel包含disk, 则内存存放不下的数据会溢写至disk, 加大延迟）。Spark 1.5以前版本，用户如果要限制Receiver的数据接收速率，可以通过设置静态配制参数“spark.streaming.receiver.maxRate”的值来实现，此举虽然可以通过限制接收速率，来适配当前的处理能力，防止内存溢出，但也会引入其它问题。比如：producer数据生产高于maxRate，当前集群处理能力也高于maxRate，这就会造成资源利用率下降等问题。为了更好的协调数据接收速率与资源处理能力，Spark Streaming 从v1.5开始引入反压机制（back-pressure）,通过动态控制数据接收速率来适配集群数据处理能力。

Backpressure

Spark Streaming Backpressure:  根据JobScheduler反馈作业的执行信息来动态调整Receiver数据接收率。通过属性spark.streaming.backpressure.enabled来控制是否启用backpressure机制，默认值false，即不启用。

• Streaming架构如下图所示（详见Streaming数据接收过程文档和Streaming 源码解析）

• BackPressure执行过程如下图所示:

在原架构的基础上加上一个新的组件RateController,这个组件负责监听“OnBatchCompleted”事件，然后从中抽取processingDelay 及schedulingDelay信息.  Estimator依据这些信息估算出最大处理速度（rate），最后由基于Receiver的Input Stream将rate通过ReceiverTracker与ReceiverSupervisorImpl转发给BlockGenerator（继承自RateLimiter）

BackPressure 源码解析

• RateController类体系

RatenController 继承自StreamingListener. 用于处理BatchCompleted事件。核心代码为：

**
* A StreamingListener that receives batch completion updates, and maintains
* an estimate of the speed at which this stream should ingest messages,
* given an estimate computation from a `RateEstimator`
*/
private[streaming] abstract class RateController(val streamUID: Int, rateEstimator: RateEstimator)
extends StreamingListener with Serializable {
……
……
private def computeAndPublish(time: Long, elems: Long, workDelay: Long, waitDelay: Long): Unit =
Future[Unit] {
val newRate = rateEstimator.compute(time, elems, workDelay, waitDelay)
newRate.foreach { s =>
rateLimit.set(s.toLong)
publish(getLatestRate())
}
}
def getLatestRate(): Long = rateLimit.get()

override def onBatchCompleted(batchCompleted: StreamingListenerBatchCompleted) {
val elements = batchCompleted.batchInfo.streamIdToInputInfo
for {
processingEnd <- batchCompleted.batchInfo.processingEndTime
workDelay <- batchCompleted.batchInfo.processingDelay
waitDelay <- batchCompleted.batchInfo.schedulingDelay
elems <- elements.get(streamUID).map(_.numRecords)
} computeAndPublish(processingEnd, elems, workDelay, waitDelay)
}
}

 

• RateController的注册

JobScheduler启动时会抽取在DStreamGraph中注册的所有InputDstream中的rateController，并向ListenerBus注册监听. 此部分代码如下：

def start(): Unit = synchronized {
if (eventLoop != null) return

logDebug("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()


for {
inputDStream <- ssc.graph.getInputStreams
rateController <- inputDStream.rateController
} ssc.addStreamingListener(rateController)

listenerBus.start()
receiverTracker = new ReceiverTracker(ssc)
inputInfoTracker = new InputInfoTracker(ssc)
receiverTracker.start()
jobGenerator.start()
logInfo("Started JobScheduler")
}

• BackPressure执行过程分析

BackPressure 执行过程分为BatchCompleted事件触发时机和事件处理两个过程

        

BatchCompleted触发过程

对BatchedCompleted的分析，应该从JobGenerator入手，因为BatchedCompleted是批次处理结束的标志，也就是JobGenerator产生的作业执行完成时触发的，因此进行作业执行分析。Streaming 应用中JobGenerator每个Batch Interval都会为应用中的每个Output Stream建立一个Job, 该批次中的所有Job组成一个Job Set.使用JobScheduler的submitJobSet进行批量Job提交。此部分代码结构如下所示

/** Generate jobs and perform checkpoint for the given `time`. */
private def generateJobs(time: Time) {
// Set the SparkEnv in this thread, so that job generation code can access the environment
// Example: BlockRDDs are created in this thread, and it needs to access BlockManager
// Update: This is probably redundant after threadlocal stuff in SparkEnv has been removed.
SparkEnv.set(ssc.env)

// Checkpoint all RDDs marked for checkpointing to ensure their lineages are
// truncated periodically. Otherwise, we may run into stack overflows (SPARK-6847).
ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")
Try {
jobScheduler.receiverTracker.allocateBlocksToBatch(time) // allocate received blocks to batch
graph.generateJobs(time) // generate jobs using allocated block
} match {
case Success(jobs) =>
val streamIdToInputInfos = jobScheduler.inputInfoTracker.getInfo(time)
jobScheduler.submitJobSet(JobSet(time, jobs, streamIdToInputInfos))
case Failure(e) =>
jobScheduler.reportError("Error generating jobs for time " + time, e)
}
eventLoop.post(DoCheckpoint(time, clearCheckpointDataLater = false))
}

其中，sumitJobSet会创建固定数量的后台线程（具体由spark.streaming.concurrentJobs指定），去处理Job Set中的Job. 具体实现逻辑为：

def submitJobSet(jobSet: JobSet) {
if (jobSet.jobs.isEmpty) {
logInfo("No jobs added for time " + jobSet.time)
} else {
listenerBus.post(StreamingListenerBatchSubmitted(jobSet.toBatchInfo))
jobSets.put(jobSet.time, jobSet)
jobSet.jobs.foreach(job => jobExecutor.execute(new JobHandler(job)))
logInfo("Added jobs for time " + jobSet.time)
}
}

其中JobHandler用于执行Job及处理Job执行结果信息。当Job执行完成时会产生JobCompleted事件. JobHandler的具体逻辑如下面代码所示：

private class JobHandler(job: Job) extends Runnable with Logging {
import JobScheduler._

def run() {
try {
val formattedTime = UIUtils.formatBatchTime(
job.time.milliseconds, ssc.graph.batchDuration.milliseconds, showYYYYMMSS = false)
val batchUrl = s"/streaming/batch/?id=${job.time.milliseconds}"
val batchLinkText = s"[output operation ${job.outputOpId}, batch time ${formattedTime}]"

ssc.sc.setJobDescription(
s"""Streaming job from <a href="$batchUrl">$batchLinkText</a>""")
ssc.sc.setLocalProperty(BATCH_TIME_PROPERTY_KEY, job.time.milliseconds.toString)
ssc.sc.setLocalProperty(OUTPUT_OP_ID_PROPERTY_KEY, job.outputOpId.toString)


ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")




var _eventLoop = eventLoop
if (_eventLoop != null) {
_eventLoop.post(JobStarted(job, clock.getTimeMillis()))



PairRDDFunctions.disableOutputSpecValidation.withValue(true) {
job.run()
}
_eventLoop = eventLoop
if (_eventLoop != null) {
_eventLoop.post(JobCompleted(job, clock.getTimeMillis()))
}
} else {

}
} finally {
ssc.sc.setLocalProperty(JobScheduler.BATCH_TIME_PROPERTY_KEY, null)
ssc.sc.setLocalProperty(JobScheduler.OUTPUT_OP_ID_PROPERTY_KEY, null)
}
}
}
}

当Job执行完成时，向eventLoop发送JobCompleted事件。EventLoop事件处理器接到JobCompleted事件后将调用handleJobCompletion 来处理Job完成事件。handleJobCompletion使用Job执行信息创建StreamingListenerBatchCompleted事件并通过StreamingListenerBus向监听器发送。实现如下：

private def handleJobCompletion(job: Job, completedTime: Long) {
val jobSet = jobSets.get(job.time)
jobSet.handleJobCompletion(job)
job.setEndTime(completedTime)
listenerBus.post(StreamingListenerOutputOperationCompleted(job.toOutputOperationInfo))
logInfo("Finished job " + job.id + " from job set of time " + jobSet.time)
if (jobSet.hasCompleted) {
jobSets.remove(jobSet.time)
jobGenerator.onBatchCompletion(jobSet.time)
logInfo("Total delay: %.3f s for time %s (execution: %.3f s)".format(
jobSet.totalDelay / 1000.0, jobSet.time.toString,
jobSet.processingDelay / 1000.0
))
listenerBus.post(StreamingListenerBatchCompleted(jobSet.toBatchInfo))
}
job.result match {
case Failure(e) =>
reportError("Error running job " + job, e)
case _ =>
}
}

BatchCompleted事件处理过程

 

StreamingListenerBus将事件转交给具体的StreamingListener，因此BatchCompleted将交由RateController进行处理。RateController接到BatchCompleted事件后将调用onBatchCompleted对事件进行处理。

override def onBatchCompleted(batchCompleted: StreamingListenerBatchCompleted) {
val elements = batchCompleted.batchInfo.streamIdToInputInfo

for {
processingEnd <- batchCompleted.batchInfo.processingEndTime
workDelay <- batchCompleted.batchInfo.processingDelay
waitDelay <- batchCompleted.batchInfo.schedulingDelay
elems <- elements.get(streamUID).map(_.numRecords)
} computeAndPublish(processingEnd, elems, workDelay, waitDelay)
}

onBatchCompleted会从完成的任务中抽取任务的执行延迟和调度延迟，然后用这两个参数用RateEstimator（目前存在唯一实现PIDRateEstimator，proportional-integral-derivative (PID) controller， PID控制器）估算出新的rate并发布。代码如下：

private def computeAndPublish(time: Long, elems: Long, workDelay: Long, waitDelay: Long): Unit =
Future[Unit] {
val newRate = rateEstimator.compute(time, elems, workDelay, waitDelay)
newRate.foreach { s =>
rateLimit.set(s.toLong)
publish(getLatestRate())
}
}

其中publish（）由RateController的子类ReceiverRateController来定义。具体逻辑如下（ReceiverInputDStream中定义）：

private[streaming] class ReceiverRateController(id: Int, estimator: RateEstimator)
extends RateController(id, estimator) {
override def publish(rate: Long): Unit =
ssc.scheduler.receiverTracker.sendRateUpdate(id, rate)
}

publish的功能为新生成的rate 借助ReceiverTracker进行转发。ReceiverTracker将rate包装成UpdateReceiverRateLimit事交ReceiverTrackerEndpoint

def sendRateUpdate(streamUID: Int, newRate: Long): Unit = synchronized {
if (isTrackerStarted) {
endpoint.send(UpdateReceiverRateLimit(streamUID, newRate))
}
}

ReceiverTrackerEndpoint接到消息后，其将会从receiverTrackingInfos列表中获取Receiver注册时使用的endpoint(实为ReceiverSupervisorImpl)，再将rate包装成UpdateLimit发送至endpoint.其接到信息后，使用updateRate更新BlockGenerators(RateLimiter子类),来计算出一个固定的令牌间隔。

private val endpoint = env.rpcEnv.setupEndpoint(
"Receiver-" + streamId + "-" + System.currentTimeMillis(), new ThreadSafeRpcEndpoint {
override val rpcEnv: RpcEnv = env.rpcEnv

override def receive: PartialFunction[Any, Unit] = {
case StopReceiver =>
logInfo("Received stop signal")
ReceiverSupervisorImpl.this.stop("Stopped by driver", None)
case CleanupOldBlocks(threshTime) =>
logDebug("Received delete old batch signal")
cleanupOldBlocks(threshTime)
case UpdateRateLimit(eps) =>
logInfo(s"Received a new rate limit: $eps.")
registeredBlockGenerators.asScala.foreach { bg =>
bg.updateRate(eps)
}
}
})

其中RateLimiter的updateRate实现如下：

private[receiver] def updateRate(newRate: Long): Unit =
if (newRate > 0) {
if (maxRateLimit > 0) {
rateLimiter.setRate(newRate.min(maxRateLimit))
} else {
rateLimiter.setRate(newRate)
}
}

setRate的实现 如下

 

public final void setRate(double permitsPerSecond) {
Preconditions.checkArgument(permitsPerSecond > 0.0
&& !Double.isNaN(permitsPerSecond), "rate must be positive");
synchronized (mutex) {
resync(readSafeMicros());
double stableIntervalMicros = TimeUnit.SECONDS.toMicros(1L) / permitsPerSecond;
this.stableIntervalMicros = stableIntervalMicros;
doSetRate(permitsPerSecond, stableIntervalMicros);
}
}

到此，backpressure反压机制调整rate结束。

流量控制点

当Receiver开始接收数据时，会通过supervisor.pushSingle()方法将接收的数据存入currentBuffer等待BlockGenerator定时将数据取走，包装成block. 在将数据存放入currentBuffer之时，要获取许可（令牌）。如果获取到许可就可以将数据存入buffer, 否则将被阻塞，进而阻塞Receiver从数据源拉取数据。

def addData(data: Any): Unit = {
if (state == Active) {
waitToPush()
synchronized {
if (state == Active) {
currentBuffer += data
} else {
throw new SparkException(
"Cannot add data as BlockGenerator has not been started or has been stopped")
}
}
} else {
throw new SparkException(
"Cannot add data as BlockGenerator has not been started or has been stopped")
}
}

其令牌投放采用令牌桶机制进行， 原理如下图所示:

令牌桶机制： 大小固定的令牌桶可自行以恒定的速率源源不断地产生令牌。如果令牌不被消耗，或者被消耗的速度小于产生的速度，令牌就会不断地增多，直到把桶填满。后面再产生的令牌就会从桶中溢出。最后桶中可以保存的最大令牌数永远不会超过桶的大小。当进行某操作时需要令牌时会从令牌桶中取出相应的令牌数，如果获取到则继续操作，否则阻塞。用完之后不用放回。

Streaming 数据流被Receiver接收后，按行解析后存入iterator中。然后逐个存入Buffer，在存入buffer时会先获取token，如果没有token存在，则阻塞；如果获取到则将数据存入buffer.  然后等价后续生成block操作。