Spark Streaming 状态管理算子的选择比较

一、updateStateByKey

官方原话:

In every batch, Spark will apply the state update function for all existing keys, 
regardless of whether they have new data in a batch or not. 
If the update function returns None then the key-value pair will be eliminated.

也即是说它会统计全局的key的状态，就算没有数据输入，它也会在每一个批次的时候返回之前的key的状态

特点:

• 大数据量的时候不适合，尤其是key维度比较高，value状态有比较大的时候。可以用redis或者alluxio。redis适合要维护key超时删除的机制的时候使用，alluxio适合超大吞吐量。当然也可以用hbase。
• key超时删除。用updatefunc返回None值。updatefanc不管是否有已存在状态的key的新数据到来，都会被已存在状态的key调用。当然，新增的key也会调用。
• 初始状态。updateStateByKey有个重载的方法可以传入initialRDD: RDD[(K, S)]

二、mapWithState

mapWithState：也是用于全局统计key的状态，但是它如果没有数据输入，便不会返回之前的key的状态，有一点增量的感觉。效率更高，生产中建议使用

object StatefulNetworkWordCount {
  def main(args: Array[String]): Unit = {
    if (args.length < 2) {
      System.err.println("Usage: StatefulNetworkWordCount  ")
      System.exit(1)
    }

    StreamingExamples.setStreamingLogLevels()

    val sparkConf = new SparkConf().setAppName("StatefulNetworkWordCount")
    // Create the context with a 1 second batch size
    val ssc = new StreamingContext(sparkConf, Seconds(1))
    ssc.checkpoint(".")

    // Initial state RDD for mapWithState operation
    val initialRDD = ssc.sparkContext.parallelize(List(("hello", 1), ("world", 1)))

    // Create a ReceiverInputDStream on target ip:port and count the
    // words in input stream of \n delimited test (eg. generated by 'nc')
    val lines = ssc.socketTextStream(args(0), args(1).toInt)
    val words = lines.flatMap(_.split(" "))
    val wordDstream = words.map(x => (x, 1))

    // Update the cumulative count using mapWithState
    // This will give a DStream made of state (which is the cumulative count of the words)
    val mappingFunc = (word: String, one: Option[Int], state: State[Int]) => {
      val sum = one.getOrElse(0) + state.getOption.getOrElse(0)
      val output = (word, sum)
      state.update(sum)
      output
    }

    val stateDstream = wordDstream.mapWithState(
      StateSpec.function(mappingFunc).initialState(initialRDD))
    stateDstream.print()
    ssc.start()
    ssc.awaitTermination()
  }
}

特点:

• 若要清除某个key的状态，可在自定义的方法中调用state.remove()
• 若要设置状态超时时间，可以调用StateSpec.function(mappingFunc).timeout()方法设置
• 若要添加初始化的状态，可以调用StateSpec.function(mappingFunc).initialState(initialRDD)方法
• 性能比updateStateByKey好

三、源码分析

updateStateByKey：

• 跟进 updateStateByKey . 其中updateFunc就是要传入的参数，他是一个函数，Seq[V]表示当前key对应的所有值，Option[S] 是当前key的历史状态，返回的是新的状态。

def updateStateByKey[S: ClassTag](
      updateFunc: (Seq[V], Option[S]) => Option[S]
    ): DStream[(K, S)] = ssc.withScope {
    updateStateByKey(updateFunc, defaultPartitioner())
  }

最终调用:

def updateStateByKey[S: ClassTag](
      updateFunc: (Iterator[(K, Seq[V], Option[S])]) => Iterator[(K, S)],
      partitioner: Partitioner,
      rememberPartitioner: Boolean): DStream[(K, S)] = ssc.withScope {
    val cleanedFunc = ssc.sc.clean(updateFunc)
    val newUpdateFunc = (_: Time, it: Iterator[(K, Seq[V], Option[S])]) => {
      cleanedFunc(it)
    }
    new StateDStream(self, newUpdateFunc, partitioner, rememberPartitioner, None)
  }

再跟进去 new StateDStream:
在这里面new出了一个StateDStream对象。在其compute方法中，会先获取上一个batch计算出的RDD（包含了至程序开始到上一个batch单词的累计计数），然后在获取本次batch中StateDStream的父类计算出的RDD（本次batch的单词计数）分别是prevStateRDD和parentRDD，然后在调用 computeUsingPreviousRDD 方法：

private [this] def computeUsingPreviousRDD(
      batchTime: Time,
      parentRDD: RDD[(K, V)],
      prevStateRDD: RDD[(K, S)]) = {
    // Define the function for the mapPartition operation on cogrouped RDD;
    // first map the cogrouped tuple to tuples of required type,
    // and then apply the update function
    val updateFuncLocal = updateFunc
    val finalFunc = (iterator: Iterator[(K, (Iterable[V], Iterable[S]))]) => {
      val i = iterator.map { t =>
        val itr = t._2._2.iterator
        val headOption = if (itr.hasNext) Some(itr.next()) else None
        (t._1, t._2._1.toSeq, headOption)
      }
      updateFuncLocal(batchTime, i)
    }
    val cogroupedRDD = parentRDD.cogroup(prevStateRDD, partitioner)
    val stateRDD = cogroupedRDD.mapPartitions(finalFunc, preservePartitioning)
    Some(stateRDD)
  }

在这里两个RDD进行cogroup然后应用updateStateByKey传入的函数。我们知道cogroup的性能是比较低下

mapWithState：

@Experimental
  def mapWithState[StateType: ClassTag, MappedType: ClassTag](
      spec: StateSpec[K, V, StateType, MappedType]
    ): MapWithStateDStream[K, V, StateType, MappedType] = {
    new MapWithStateDStreamImpl[K, V, StateType, MappedType](
      self,
      spec.asInstanceOf[StateSpecImpl[K, V, StateType, MappedType]]
    )
  }

说明：StateSpec 封装了状态管理函数，并在该方法中创建了MapWithStateDStreamImpl对象。
MapWithStateDStreamImpl 中创建了一个InternalMapWithStateDStream类型对象internalStream，在MapWithStateDStreamImpl的compute方法中调用了internalStream的getOrCompute方法。
MapWithStateDStreamImpl 中创建了一个InternalMapWithStateDStream类型对象internalStream，在MapWithStateDStreamImpl的compute方法中调用了internalStream的getOrCompute方法。

private[streaming] class MapWithStateDStreamImpl[
    KeyType: ClassTag, ValueType: ClassTag, StateType: ClassTag, MappedType: ClassTag](
    dataStream: DStream[(KeyType, ValueType)],
    spec: StateSpecImpl[KeyType, ValueType, StateType, MappedType])
  extends MapWithStateDStream[KeyType, ValueType, StateType, MappedType](dataStream.context) {

  private val internalStream =
    new InternalMapWithStateDStream[KeyType, ValueType, StateType, MappedType](dataStream, spec)

  override def slideDuration: Duration = internalStream.slideDuration

  override def dependencies: List[DStream[_]] = List(internalStream)

  override def compute(validTime: Time): Option[RDD[MappedType]] = {
    internalStream.getOrCompute(validTime).map { _.flatMap[MappedType] { _.mappedData } }
  }

InternalMapWithStateDStream中没有getOrCompute方法，这里调用的是其父类 DStream 的getOrCpmpute方法，该方法中最终会调用InternalMapWithStateDStream的Compute方法：

/** Method that generates an RDD for the given time */
  override def compute(validTime: Time): Option[RDD[MapWithStateRDDRecord[K, S, E]]] = {
    // Get the previous state or create a new empty state RDD
    val prevStateRDD = getOrCompute(validTime - slideDuration) match {
      case Some(rdd) =>
        if (rdd.partitioner != Some(partitioner)) {
          // If the RDD is not partitioned the right way, let us repartition it using the
          // partition index as the key. This is to ensure that state RDD is always partitioned
          // before creating another state RDD using it
          MapWithStateRDD.createFromRDD[K, V, S, E](
            rdd.flatMap { _.stateMap.getAll() }, partitioner, validTime)
        } else {
          rdd
        }
      case None =>
        MapWithStateRDD.createFromPairRDD[K, V, S, E](
          spec.getInitialStateRDD().getOrElse(new EmptyRDD[(K, S)](ssc.sparkContext)),
          partitioner,
          validTime
        )
    }


    // Compute the new state RDD with previous state RDD and partitioned data RDD
    // Even if there is no data RDD, use an empty one to create a new state RDD
    val dataRDD = parent.getOrCompute(validTime).getOrElse {
      context.sparkContext.emptyRDD[(K, V)]
    }
    val partitionedDataRDD = dataRDD.partitionBy(partitioner)
    val timeoutThresholdTime = spec.getTimeoutInterval().map { interval =>
      (validTime - interval).milliseconds
    }
    Some(new MapWithStateRDD(
      prevStateRDD, partitionedDataRDD, mappingFunction, validTime, timeoutThresholdTime))
  }

根据给定的时间生成一个MapWithStateRDD，首先获取了先前状态的RDD：preStateRDD和当前时间的RDD:dataRDD，然后对dataRDD基于先前状态RDD的分区器进行重新分区获取partitionedDataRDD。并且保证这两个RDD使用的是同样的partitioner。MapWithStateRDD中的数据都是MapWithStateRDDRecord对象，每个分区对应一个对象来保存状态（这就是为什么两个RDD需要用同一个Partitioner）。最后将preStateRDD，partitionedDataRDD和用户定义的函数mappingFunction传给新生成的MapWithStateRDD对象返回。

跟进MapWithStateRDD的compute方法:

override def compute(
      partition: Partition, context: TaskContext): Iterator[MapWithStateRDDRecord[K, S, E]] = {

    val stateRDDPartition = partition.asInstanceOf[MapWithStateRDDPartition]
    val prevStateRDDIterator = prevStateRDD.iterator(
      stateRDDPartition.previousSessionRDDPartition, context)
    val dataIterator = partitionedDataRDD.iterator(
      stateRDDPartition.partitionedDataRDDPartition, context)

    val prevRecord = if (prevStateRDDIterator.hasNext) Some(prevStateRDDIterator.next()) else None
    val newRecord = MapWithStateRDDRecord.updateRecordWithData(
      prevRecord,
      dataIterator,
      mappingFunction,
      batchTime,
      timeoutThresholdTime,
      removeTimedoutData = doFullScan // remove timed-out data only when full scan is enabled
    )
    Iterator(newRecord)
  }

首先获取了先前状态的RDD:preStateRDD和当前时间的RDD:dataRDD的迭代器，接着获取了prevStateRDD的一条数据，这个分区也只有一条MapWithStateRDDRecord类型的数据，维护了对应分区所有数据状态。
然后调用MapWithStateRDDRecord的updateRecordWithData方法,用当前时间的RDD数据更新之前的状态RDD数据.

private[streaming] object MapWithStateRDDRecord {
  def updateRecordWithData[K: ClassTag, V: ClassTag, S: ClassTag, E: ClassTag](
    prevRecord: Option[MapWithStateRDDRecord[K, S, E]],
    dataIterator: Iterator[(K, V)],
    mappingFunction: (Time, K, Option[V], State[S]) => Option[E],
    batchTime: Time,
    timeoutThresholdTime: Option[Long],
    removeTimedoutData: Boolean
  ): MapWithStateRDDRecord[K, S, E] = {
    // Create a new state map by cloning the previous one (if it exists) or by creating an empty one
    val newStateMap = prevRecord.map { _.stateMap.copy() }. getOrElse { new EmptyStateMap[K, S]() }

    val mappedData = new ArrayBuffer[E]
    val wrappedState = new StateImpl[S]()

    // Call the mapping function on each record in the data iterator, and accordingly
    // update the states touched, and collect the data returned by the mapping function
    dataIterator.foreach { case (key, value) =>
      wrappedState.wrap(newStateMap.get(key))
      val returned = mappingFunction(batchTime, key, Some(value), wrappedState)
      if (wrappedState.isRemoved) {
        newStateMap.remove(key)
      } else if (wrappedState.isUpdated
          || (wrappedState.exists && timeoutThresholdTime.isDefined)) {
        newStateMap.put(key, wrappedState.get(), batchTime.milliseconds)
      }
      mappedData ++= returned
    }

    // Get the timed out state records, call the mapping function on each and collect the
    // data returned
    if (removeTimedoutData && timeoutThresholdTime.isDefined) {
      newStateMap.getByTime(timeoutThresholdTime.get).foreach { case (key, state, _) =>
        wrappedState.wrapTimingOutState(state)
        val returned = mappingFunction(batchTime, key, None, wrappedState)
        mappedData ++= returned
        newStateMap.remove(key)
      }
    }

    MapWithStateRDDRecord(newStateMap, mappedData)
  }
}

先copy了原来的状态，接着定义了两个变量，mappedData是最终要返回的结果，wrappedState可以看成是对state的包装，添加了一些额外的方法。
接着遍历当前批次的数据，从状态中取出key对应的原来的state，并根据自定义的函数来对state进行更新，这里涉及到state的remove&update&timeout来对newStateMap进行更新操作，并将有更新的状态加入到了mappedData中。
若有设置超时时间，则还会对超时了的key进行移除，也会加入到mappedData中，最终通过新的状态对象newStateMap和需返回的mappedData数组构建了MapWithStateRDDRecord对象来返回。

四、总结

updateStateByKey底层是将preSateRDD和parentRDD进行co-group，然后对所有数据都将经过自定义的mapFun函数进行一次计算，即使当前batch只有一条数据也会进行这么复杂的计算，大大的降低了性能，并且计算时间会随着维护的状态的增加而增加。

mapWithstate底层是创建了一个MapWithStateRDD，存的数据是MapWithStateRDDRecord对象，一个Partition对应一个MapWithStateRDDRecord对象，该对象记录了对应Partition所有的状态，每次只会对当前batch有的数据进行跟新，而不会像updateStateByKey一样对所有数据计算。