本篇博客将详细探讨DStream模板下的RDD是如何被创建,然后被执行的。在开始叙述之前,先来思考几个问题,本篇文章也就是基于此问题构建的。 
1. RDD是谁产生的? 
2. 如何产生RDD? 
带着这两个问题开启我们的探索之旅。

DStream是RDD的模板,每隔一个Batch Interval会根据DStream模板生成一个对应的RDD,然后将RDD存储到DStream中的generatedRDDs数据结构中,下面是存储结构格式。

// RDDs generated, marked as private[streaming] so that testsuites can access it
@transient
private[streaming] var generatedRDDs = new HashMap[Time, RDD[T]] ()

1、简单的WordCount程序

object WordCount {  def main(args:Array[String]): Unit ={
   val sparkConf = new SparkConf().setMaster("Master:7077").setAppName("WordCount")
   val ssc = new StreamingContext(sparkConf,Seconds(10)) // Timer触发频率

   val lines = ssc.socketTextStream("Master",9999) //接收数据
   val words = lines.flatMap(_.split(" "))
   val wordCounts = words.map(x => (x,1)).reduceByKey(_+_)
   wordCounts.print()
   ssc.start()
   ssc.awaitTermination()
 }
}
首先我们先看看print方法,具体的代码如下:

/**
* Print the first num elements of each RDD generated in this DStream. This is an output
* operator, so this DStream will be registered as an output stream and there materialized.
*/
def print(num: Int): Unit = ssc.withScope {
 def foreachFunc: (RDD[T], Time) => Unit = {
   (rdd: RDD[T], time: Time) => {
     val firstNum = rdd.take(num + 1)
     // scalastyle:off println
     
println("-------------------------------------------")
     println("Time: " + time)
     println("-------------------------------------------")
     firstNum.take(num).foreach(println)
     if (firstNum.length > num) println("...")
     println()
     // scalastyle:on println
   
}
 }
 
foreachRDD(context.sparkContext.clean(foreachFunc), displayInnerRDDOps = false)
}

首先定义了一个函数,该函数用来从RDD中取出前几条数据,并打印出结果与时间等,后面会调用foreachRDD函数。

private def foreachRDD(
   foreachFunc: (RDD[
T], Time) => Unit,
   
displayInnerRDDOps: Boolean): Unit = {
   
new ForEachDStream(this,context.sparkContext.clean(foreachFunc, false), displayInnerRDDOps).register()
}

/**
* Register this streaming as an output stream. This would ensure that RDDs of this
* DStream will be generated.
*/
private[streaming] def register(): DStream[T] = {
 ssc.graph.addOutputStream(this)
 this
}

def addOutputStream(outputStream: DStream[_]) {
 this.synchronized {
   outputStream.setGraph(this)
   outputStreams += outputStream
 }

在foreachRDD中new出了一个ForEachDStream对象,并将这个注册给DStreamGraph,ForEachDStream对象也就是DStreamGraph中的outputStreams。

当每到达一个BatchInterval时候,就会调用DStreamingGraph中的generateJobs.


def generateJobs(time: Time): Seq[Job] = {
 logDebug(
"Generating jobs for time " + time)
 
val jobs = this.synchronized {
   
outputStreams.flatMap { outputStream =>
     
val jobOption = outputStream.generateJob(time)
     jobOption.foreach(_.setCallSite(outputStream.
creationSite))
     jobOption
   }
 }
 logDebug(
"Generated " + jobs.length + " jobs for time " + time)
 jobs
}

这里就会调用outputStream的generateJob方法


private[streaming] def generateJob(time: Time): Option[Job] = {
 
getOrCompute(time) match {
   
case Some(rdd) => {
     
val jobFunc = () => {
       
val emptyFunc = { (iterator: Iterator[T]) => {} }
       context.sparkContext.runJob(rdd
, emptyFunc)
     }
     
Some(new Job(time, jobFunc))
   }
   
case None => None
 }
}

这里会调用getOrCompute(time)来产生新RDD,并将其存入到generatedRDDs中,整理的过程如下图:
(版本定制)第8课:Spark Streaming源码解读之RDD生成生命周期彻底研究和思考_streaming

参考博客:

备注:

1、DT大数据梦工厂微信公众号DT_Spark 
2、IMF晚8点大数据实战YY直播频道号:68917580
3、新浪微博: http://www.weibo.com/ilovepains