Spark Streaming发行版笔记14：updateStateByKey和mapWithState源码解密

Posted 2020-07-17 永不服输2016

本篇从二个方面进行源码分析：

一、updateStateByKey解密

二、mapWithState解密

 

通过对Spark研究角度来研究jvm、分布式、图计算、架构设计、软件工程思想，可以学到很多东西。

进行黑名单动态生成和过滤例子中会用到updateStateByKey方法，此方法在DStream类中没有定义，需要在

DStream的object区域通过隐式转换来找，如下面的代码：

object DStream {
  // `toPairDStreamFunctions` was in SparkContext before 1.3 and users had to
  // `import StreamingContext._` to enable it. Now we move it here to make the compiler find
  // it automatically. However, we still keep the old function in StreamingContext for backward
  // compatibility and forward to the following function directly.
  implicit def toPairDStreamFunctions[K, V](stream: DStream[(K, V)])
      (implicit kt: ClassTag[K], vt: ClassTag[V], ord: Ordering[K] = null):
    PairDStreamFunctions[K, V] = {
    new PairDStreamFunctions[K, V](stream)
  }

继续跟踪PairDStreamFunctions类中有次方法定义：

/**
 * Return a new "state" DStream where the state for each key is updated by applying
 * the given function on the previous state of the key and the new values of each key.
 * Hash partitioning is used to generate the RDDs with `numPartitions` partitions.
 * @param updateFunc State update function. If `this` function returns None, then
 * corresponding state key-value pair will be eliminated.
 * @param numPartitions Number of partitions of each RDD in the new DStream.
 * @tparam S State type
 */

def updateStateByKey[S: ClassTag](
    updateFunc: (Seq[V], Option[S]) => Option[S],
    numPartitions: Int
  ): DStream[(K, S)] = ssc.withScope {
  updateStateByKey(updateFunc, defaultPartitioner(numPartitions))
}
继续返回DStream类：
HashPartitioner的特点是效率高，spark1.2之前采用的主要目的是效率高，不需要排序之类的，设置并行度：
private[streaming] def defaultPartitioner(numPartitions: Int = self.ssc.sc.defaultParallelism) = {
  new HashPartitioner(numPartitions)
}
/**

 * Return a new "state" DStream where the state for each key is updated by applying
 * the given function on the previous state of the key and the new values of each key.
 * org.apache.spark.Partitioner is used to control the partitioning of each RDD.
 * @param updateFunc State update function. Note, that this function may generate a different
 * tuple with a different key than the input key. Therefore keys may be removed
 * or added in this way. It is up to the developer to decide whether to
 * remember the partitioner despite the key being changed.
 * @param partitioner Partitioner for controlling the partitioning of each RDD in the new
 * DStream
 * @param rememberPartitioner Whether to remember the paritioner object in the generated RDDs.
 * @tparam S State type
 */

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

继续跟踪StateDStream，继承了DStream，如果对状态不断的操作就会产生很多的StateDStream状态对象：

private[streaming]
class StateDStream[K: ClassTag, V: ClassTag, S: ClassTag](
    parent: DStream[(K, V)],
    updateFunc: (Iterator[(K, Seq[V], Option[S])]) => Iterator[(K, S)],
    partitioner: Partitioner,
    preservePartitioning: Boolean,
    initialRDD : Option[RDD[(K, S)]]
  ) extends DStream[(K, S)](parent.ssc) {
  super.persist(StorageLevel.MEMORY_ONLY_SER)

看一段关键的代码：

override def compute(validTime: Time): Option[RDD[(K, S)]] = {
  // Try to get the previous state RDD
  getOrCompute(validTime - slideDuration) match {
    case Some(prevStateRDD) => {    // If previous state RDD exists
      // Try to get the parent RDD
      parent.getOrCompute(validTime) match {
        case Some(parentRDD) => {   // If parent RDD exists, then compute as usual
          computeUsingPreviousRDD (parentRDD, prevStateRDD)
        }
        case None => {    // If parent RDD does not exist
          // Re-apply the update function to the old state RDD
          val updateFuncLocal = updateFunc
          val finalFunc = (iterator: Iterator[(K, S)]) => {
            val i = iterator.map(t => (t._1, Seq[V](), Option(t._2)))
            updateFuncLocal(i)
          }//效率角度

          val stateRDD = prevStateRDD.mapPartitions(finalFunc, preservePartitioning)
          Some(stateRDD)
        }
      }
    }
根据代码分析，把函数传进来，看cogroup，按照key对value进行聚合，按照key对所有数据进行扫描然后聚合，这样做好处是对rdd的计算；

不好的地方就是性能，cogroup对所有数据进行扫描，随着时间流逝数据规模越来越大性能越低，cogroup rdd和另一个

cogroup rdd数据进行扫描合并。如下关键代码:

private [this] def computeUsingPreviousRDD (
  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(i)
  }
  val cogroupedRDD = parentRDD.cogroup(prevStateRDD, partitioner)
  val stateRDD = cogroupedRDD.mapPartitions(finalFunc, preservePartitioning)
  Some(stateRDD)
}

/**
 * For each key k in `this` or `other`, return a resulting RDD that contains a tuple with the
 * list of values for that key in `this` as well as `other`.
 */
def cogroup[W](other: RDD[(K, W)], partitioner: Partitioner)
    : RDD[(K, (Iterable[V], Iterable[W]))] = self.withScope {
  if (partitioner.isInstanceOf[HashPartitioner] && keyClass.isArray) {
    throw new SparkException("Default partitioner cannot partition array keys.")
  }
  val cg = new CoGroupedRDD[K](Seq(self, other), partitioner)
  cg.mapValues { case Array(vs, w1s) =>
    (vs.asInstanceOf[Iterable[V]], w1s.asInstanceOf[Iterable[W]])
  }
}

继续剖析mapWithState解密

再看mapWithState，返回的是一个DStream，维护历史状态、更新历史状态都是基于key来维护，state相当于内存数据表，其实是在删除一张表，这张表中

记录了历史状态，一张key、value、state的表，所有历史状态都放在这张表中，根据key 在satate的基础上更新value，如单词计数，不断累积计数：

/**
 * :: Experimental ::
 * Return a [[MapWithStateDStream]] by applying a function to every key-value element of
 * `this` stream, while maintaining some state data for each unique key. The mapping function
 * and other specification (e.g. partitioners, timeouts, initial state data, etc.) of this
 * transformation can be specified using [[StateSpec]] class. The state data is accessible in
 * as a parameter of type [[State]] in the mapping function.
 *
 * Example of using `mapWithState`:
 * {{{
 *    // A mapping function that maintains an integer state and return a String
 *    def mappingFunction(key: String, value: Option[Int], state: State[Int]): Option[String] = {
 *      // Use state.exists(), state.get(), state.update() and state.remove()
 *      // to manage state, and return the necessary string
 *    }
 *
 *    val spec = StateSpec.function(mappingFunction).numPartitions(10)
 *
 *    val mapWithStateDStream = keyValueDStream.mapWithState[StateType, MappedType](spec)
 * }}}
 *
 * @param spec          Specification of this transformation
 * @tparam StateType    Class type of the state data
 * @tparam MappedType   Class type of the mapped data
 */
@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]]
  )
}

内存数据表都会有：defined、timingOut、updated、removed：
/**

 * :: Experimental ::
 * Abstract class for getting and updating the state in mapping function used in the `mapWithState`
 * operation of a [[org.apache.spark.streaming.dstream.PairDStreamFunctions pair DStream]] (Scala)
 * or a [[org.apache.spark.streaming.api.java.JavaPairDStream JavaPairDStream]] (Java).
 *
 * Scala example of using `State`:
 * {{{
 *    // A mapping function that maintains an integer state and returns a String
 *    def mappingFunction(key: String, value: Option[Int], state: State[Int]): Option[String] = {
 *      // Check if state exists
 *      if (state.exists) {
 *        val existingState = state.get  // Get the existing state
 *        val shouldRemove = ...         // Decide whether to remove the state
 *        if (shouldRemove) {
 *          state.remove()     // Remove the state
 *        } else {
 *          val newState = ...
 *          state.update(newState)    // Set the new state
 *        }
 *      } else {

 *        val initialState = ...
 *        state.update(initialState)  // Set the initial state
 *      }
 *      ... // return something
 *    }
 *
 * }}}

/** Internal implementation of the [[State]] interface */
private[streaming] class StateImpl[S] extends State[S] {
  private var state: S = null.asInstanceOf[S]
  private var defined: Boolean = false
  private var timingOut: Boolean = false
  private var updated: Boolean = false
  private var removed: Boolean = false
  // ========= Public API =========
  override def exists(): Boolean = {
    defined
  }
  override def get(): S = {
    if (defined) {
      state
    } else {
      throw new NoSuchElementException("State is not set")
    }
  }
  override def update(newState: S): Unit = {
    require(!removed, "Cannot update the state after it has been removed")
    require(!timingOut, "Cannot update the state that is timing out")
    state = newState
    defined = true
    updated = true
  }

下面的代码V就是外面传入的函数：

/** Internal implementation of [[org.apache.spark.streaming.StateSpec]] interface. */
private[streaming]
case class StateSpecImpl[K, V, S, T](
 function: (Time, K, Option[V], State[S]) => Option[T]) extends StateSpec[K, V, S, T] {
 /** Internal implementation of the [[MapWithStateDStream]] */

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 } }
  }

基于历史数据的更新，有内存数据结构，更新已有数据结构，而不是在已有的基础上创建内存数据结构：

/**
 * A DStream that allows per-key state to be maintains, and arbitrary records to be generated
 * based on updates to the state. This is the main DStream that implements the `mapWithState`
 * operation on DStreams.
 *
 * @param parent Parent (key, value) stream that is the source
 * @param spec Specifications of the mapWithState operation
 * @tparam K   Key type
 * @tparam V   Value type
 * @tparam S   Type of the state maintained
 * @tparam E   Type of the mapped data
 */

private[streaming]
class InternalMapWithStateDStream[K: ClassTag, V: ClassTag, S: ClassTag, E: ClassTag](
    parent: DStream[(K, V)], spec: StateSpecImpl[K, V, S, E])
  extends DStream[MapWithStateRDDRecord[K, S, E]](parent.context) {
  persist(StorageLevel.MEMORY_ONLY)

基于时间窗口创建一个新rdd，是所有故事下面开始：

Some(new MapWithStateRDD(
  prevStateRDD, partitionedDataRDD, mappingFunction, validTime, timeoutThresholdTime))
** Method that generates a 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))
  }
}

看下MapWithStateRDD：

/**
 * RDD storing the keyed states of `mapWithState` operation and corresponding mapped data.
 * Each partition of this RDD has a single record of type [[MapWithStateRDDRecord]]. This contains a
 * [[StateMap]] (containing the keyed-states) and the sequence of records returned by the mapping
 * function of  `mapWithState`.
 * @param prevStateRDD The previous MapWithStateRDD on whose StateMap data `this` RDD
  *                    will be created
 * @param partitionedDataRDD The partitioned data RDD which is used update the previous StateMaps
 *                           in the `prevStateRDD` to create `this` RDD
 * @param mappingFunction  The function that will be used to update state and return new data
 * @param batchTime        The time of the batch to which this RDD belongs to. Use to update
 * @param timeoutThresholdTime The time to indicate which keys are timeout
 */

private[streaming] class MapWithStateRDD[K: ClassTag, V: ClassTag, S: ClassTag, E: ClassTag](
    private var prevStateRDD: RDD[MapWithStateRDDRecord[K, S, E]],
    private var partitionedDataRDD: RDD[(K, V)],
    mappingFunction: (Time, K, Option[V], State[S]) => Option[E],
    batchTime: Time,
    timeoutThresholdTime: Option[Long]
  ) extends RDD[MapWithStateRDDRecord[K, S, E]](
    partitionedDataRDD.sparkContext,
    List(
      new OneToOneDependency[MapWithStateRDDRecord[K, S, E]](prevStateRDD),
      new OneToOneDependency(partitionedDataRDD))
  ) {

每个partition被一个MapWithStateRDDRecord代表的，里面有一个数据结构stateMap，再看此类的重点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 timedout data only when full scan is enabled
  )
  Iterator(newRecord)
}

 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)
}

wrappedState是可以不断被赋值的，mappedData代表最后返回的值。根据当前batch的数据进行计算，更新了newStateMap的数据结构，保存了历史数据，

没有对历史数据进行计算或遍历，只会进行更新、插入操作。Record代表一个partition，MapWithStateRDDRecord中record记录并没改变。

DStream操作RDD，RDD内部变了。所以不可变的rdd可以处理变化的rdd。

感谢王家林老师的知识分享

王家林老师名片：

中国Spark第一人

感谢王家林老师的知识分享

新浪微博：http://weibo.com/ilovepains

微信公众号：DT_Spark

博客：http://blog.sina.com.cn/ilovepains

手机：[phone]

QQ：1740415547

邮箱：[email protected]

YY课堂：每天20:00现场授课频道68917580