1. 云图网首页
  2. 技术专区
  3. 大数据

Spark算子执行流程详解之一大数据

• 大数据

1.take

获取前num条记录。

def take(num: Int): Array[T] = withScope {
  if (num == 0) {
    new Array[T](0)
  } else {
    val buf = newArrayBuffer[T]
    val totalParts = this.partitions.length
    var partsScanned = 0
    while (buf.size < num && partsScanned < totalParts) {
      // The number of partitions to try in this iteration. It is ok for this number to be
      // greater than totalParts because we actually cap it at totalParts in runJob.
      var numPartsToTry =1
      if (partsScanned > 0) {
        // If we didn’t find any rows after the previous iteration, quadruple and retry.
        // Otherwise, interpolate the number of partitions we need to try, but overestimate
        // it by 50%. We also cap the estimation in the end.
        if (buf.size ==0) {//截止目前为止buf为空的话，则扩大4倍范围
          numPartsToTry = partsScanned * 4
        } else {//截止目前为止还有部分值没取到的话，则扩大至Math.max((1.5 * num * partsScanned / buf.size).toInt – partsScanned, 1)，但是不超过当前已扫描过分区的4倍
          // the left side of max is >=1 whenever partsScanned >= 2
          numPartsToTry = Math.max((1.5* num * partsScanned / buf.size).toInt – partsScanned, 1)
          numPartsToTry = Math.min(numPartsToTry, partsScanned * 4)
        }
      }

      val left = num – buf.size
      val p = partsScanned until math.min(partsScanned + numPartsToTry, totalParts)
      val res = sc.runJob(this, (it:Iterator[T]) => it.take(left).toArray, p, allowLocal =true)

      res.foreach(buf ++= _.take(num – buf.size))
      partsScanned += numPartsToTry
    }

    buf.toArray
  }
}

首先关注下sc.runJob函数的传参：

/**
 * Run a job on a given set of partitions of an RDD, but take a function of type
 * `Iterator[T] => U` instead of `(TaskContext, Iterator[T]) => U`.
 */
def runJob[T,U: ClassTag](
    rdd: RDD[T],
    func: Iterator[T] =>U,
    partitions: Seq[Int],
    allowLocal: Boolean
    ): Array[U] = {
  val cleanedFunc = clean(func)
  runJob(rdd, (ctx: TaskContext, it: Iterator[T]) => cleanedFunc(it), partitions, allowLocal)
}

 

其中partitions: Seq[Int]代表需要计算的分区，可以计算某个分区，也可以计算多个分区，是待计算的分区集合。

其次先看第一次循环，其partsScanned为0，numPartsToTry为1，因此先计算第一个分区的结果，如果第一次计算可以取得满足条件的num个值，则循环结束，如果取不到满足条件的num个值，则扩大第二次计算的分区范围，很可能一下子扫多个分区。

其执行过程见下图：

Spark算子执行流程详解之一大数据

Take可以避免全量计算，执行时间比较短。但可能会多次触发action。

2.first

取RDD的第一个元素

/**
 * Return the first element in this RDD.
 */
def first(): T = withScope {
  take(1) match {
    case Array(t) => t
    case _ => throw newUnsupportedOperationException(“empty collection”)
  }

}

其实就是调用take来完成的，take的流程可以查阅take函数详解

3.sortByKey

 

def sortByKey(ascending: Boolean =true, numPartitions: Int = self.partitions.length)
    : RDD[(K, V)] = self.withScope
{
  val part = newRangePartitioner(numPartitions, self, ascending)
  new ShuffledRDD[K,V, V](self, part)
    .setKeyOrdering(if (ascending)ordering elseordering.reverse)
}

sortByKey其实就是根据父RDD生成ShuffledRDD的过程，其分区函数为范围分区RangePartitioner，执行过程如下：

Spark算子执行流程详解之一大数据

父RDD的每个分区按照分区函数RangePartitioner将每个分区的数据划分为多个分区的数据，然后ShuffledRDD拉取自己对应分区的数据。但是sortByKey主要应该掌握其RangePartitioner分区函数的执行原理，它如何保证ShuffledRDD的每个分区的数量是大致相同的，也就是如何来划分每个分区的边界的，且看：

class RangePartitioner[K: Ordering : ClassTag, V](
    @transient partitions: Int,
    @transient rdd: RDD[_ <: Product2[K,V]],
    private var ascending: Boolean =true)
  extends Partitioner {

  // We allow partitions = 0, which happens when sorting an empty RDD under the default settings.
  require(partitions >= 0, s”Number of partitions cannot be negative but found$partitions.”)

  private var ordering= implicitly[Ordering[K]]

  // An array of upper bounds for the first (partitions – 1) partitions前（partitions – 1）的分区边界
  private var rangeBounds: Array[K] = {
    if (partitions <= 1) {
      Array.empty
    } else {
      // This is the sample size we need to have roughly balanced output partitions, capped at 1M.
      val sampleSize = math.min(20.0* partitions, 1e6)
      // Assume the input partitions are roughly balanced and over-sample a little bit.
      val sampleSizePerPartition = math.ceil(3.0* sampleSize / rdd.partitions.size).toInt
      // numItems相当于记录rdd元素的总数
      // sketched的类型是Array[(Int, Int, Array[K])]，记录的是分区的编号、该分区中总元素的个数以及从父RDD中每个分区采样的数据
      val (numItems, sketched) = RangePartitioner.sketch(rdd.map(_._1), sampleSizePerPartition)
      if (numItems == 0L) {
        Array.empty
      } else {
        // If a partition contains much more than the average number of items, we re-sample from it
        // to ensure that enough items are collected from that partition.
        val fraction = math.min(sampleSize / math.max(numItems,1L), 1.0)
        val candidates = ArrayBuffer.empty[(K, Float)]
        val imbalancedPartitions = mutable.Set.empty[Int]
        sketched.foreach { case (idx,n, sample) =>
          if (fraction * n > sampleSizePerPartition) {
            imbalancedPartitions += idx
          } else {
            // The weight is 1 over the sampling probability.
            val weight = (n.toDouble / sample.size).toFloat
            for (key<- sample) {
              candidates += ((key, weight))
            }
          }
        }
        if (imbalancedPartitions.nonEmpty) {
          // Re-sample imbalanced partitions with the desired sampling probability.
          val imbalanced =new PartitionPruningRDD(rdd.map(_._1), imbalancedPartitions.contains)
          val seed = byteswap32(-rdd.id– 1)
          val reSampled = imbalanced.sample(withReplacement =false, fraction, seed).collect()
          val weight = (1.0/ fraction).toFloat
          candidates ++= reSampled.map(x => (x, weight))
        }
        RangePartitioner.determineBounds(candidates, partitions)
      }
    }
  }

  def numPartitions: Int = rangeBounds.length +1

  private var binarySearch: ((Array[K],K) => Int) = CollectionsUtils.makeBinarySearch[K]

  def getPartition(key: Any): Int = {
    val k = key.asInstanceOf[K]
    var partition = 0
    if (rangeBounds.length <=128) {
      // If we have less than 128 partitions naive search
      while (partition <rangeBounds.length && ordering.gt(k,rangeBounds(partition))) {
        partition += 1
      }
    } else {
      // Determine which binary search method to use only once.
      partition = binarySearch(rangeBounds, k)
      // binarySearch either returns the match location or -[insertion point]-1
      if (partition <0) {
        partition = -partition-1
      }
      if (partition > rangeBounds.length) {
        partition = rangeBounds.length
      }
    }
    if (ascending) {
      partition
    } else {
      rangeBounds.length – partition
    }
  }
private[spark] objectRangePartitioner {

  /**
   * Sketches the input RDD via reservoir sampling on each partition.
   *
   * @param rdd the input RDD to sketch
   * @param sampleSizePerPartition max sample size per partition
   * @return (total number of items, an array of (partitionId, number of items, sample))
   */
  def sketch[K: ClassTag](
      rdd: RDD[K],
      sampleSizePerPartition: Int): (Long, Array[(Int, Int, Array[K])]) = {
    val shift = rdd.id
    // val classTagK = classTag[K] // to avoid serializing the entire partitioner object
    val sketched = rdd.mapPartitionsWithIndex { (idx, iter) =>
      val seed = byteswap32(idx ^ (shift <<16))
      //Reservoir:水塘抽样
      val (sample, n) = SamplingUtils.reservoirSampleAndCount(
        iter, sampleSizePerPartition, seed)
      Iterator((idx, n, sample))
    }.collect()
    val numItems = sketched.map(_._2.toLong).sum
    (numItems, sketched)
  }

  /**
   * Determines the bounds for range partitioning from candidates with weights indicating how many
   * items each represents. Usually this is 1 over the probability used to sample this candidate.
   *
   * @param candidates unordered candidates with weights
   * @param partitions number of partitions
   * @return selected bounds
   */
  def determineBounds[K: Ordering : ClassTag](
      candidates: ArrayBuffer[(K, Float)],
      partitions: Int): Array[K] = {
    val ordering = implicitly[Ordering[K]]
    val ordered = candidates.sortBy(_._1)
    val numCandidates = ordered.size
    val sumWeights = ordered.map(_._2.toDouble).sum
    val step = sumWeights / partitions
    var cumWeight = 0.0
    var target = step
    val bounds = ArrayBuffer.empty[K]
    var i = 0
    var j = 0
    var previousBound = Option.empty[K]
    while ((i < numCandidates) && (j < partitions –1)) {
      val (key, weight) = ordered(i)
      cumWeight += weight
      if (cumWeight > target) {
        // Skip duplicate values.
        if (previousBound.isEmpty || ordering.gt(key, previousBound.get)) {
          bounds += key
          target += step
          j += 1
          previousBound = Some(key)
        }
      }
      i += 1
    }
    bounds.toArray
  }

}

可见其分区边界由rangeBounds保存，然后提供getPartition函数，根据传入的key获取其对于的分区编号。那么现在的问题就是当不知道父RDD的每个分区总数的情况下，如何保证数据被随机抽样出来，只有数据随机被抽样出来，才能保证之后切分分区的时候每个分区的数目是大致相同的。（这样就可以只扫描一次获取随机值，否则需要先扫描出总数，然后根据总数来抽样，这样就扫描了2次）

       首先抛出个数学算法，即 Reservoir Sampling（水塘抽样），目的在于从包含n个项目的集合S中选取k个样本，其中n为一很大或未知的数量，尤其适用于不能把所有n个项目都存放到主内存的情况。

       具体推导过程这里不详细描述，直接写结论：

在取第n个数据的时候，我们生成一个0到1的随机数p，如果p小于k/n，替换池中任意一个为第n个数。大于k/n，继续保留前面的数。直到数据流结束，返回此k个数。但是为了保证计算机计算分数额准确性，一般是生成一个0到n的随机数。

//stream代表数据流 

//reservoir代表返回长度为k的池塘 

//从stream中取前k个放入reservoir； 

for ( int i = 1; i < k; i++) 

    reservoir[i] = stream[i]; 

for (i = k; stream != null; i++) { 

    p = random(0, i); 

    if (p < k) reservoir[p] = stream[i]; 

return reservoir; 

接下来看其具体的执行过程：

private var rangeBounds: Array[K] = {
    if (partitions <= 1) {
      Array.empty
    } else {
      // This is the sample size we need to have roughly balanced output partitions, capped at 1M.
      val sampleSize = math.min(20.0* partitions, 1e6)
      // Assume the input partitions are roughly balanced and over-sample a little bit.
      val sampleSizePerPartition = math.ceil(3.0* sampleSize / rdd.partitions.size).toInt
      // numItems相当于记录rdd元素的总数
      // sketched的类型是Array[(Int, Int, Array[K])]，记录的是分区的编号、该分区中总元素的个数以及从父RDD中每个分区采样的数据

// sampleSizePerPartition代表的是每个分区抽样的值，然后针对待排序的key值进行抽样，即sketch函数
      val (numItems, sketched) = RangePartitioner.sketch(rdd.map(_._1), sampleSizePerPartition)
      if (numItems == 0L) {
        Array.empty
      } else {
        // If a partition contains much more than the average number of items, we re-sample from it
        // to ensure that enough items are collected from that partition.

//如果存在数据倾斜的情况，则某些分区包含数据量多的情况下，抽样的值偏少，需要增加抽样的数目
        val fraction = math.min(sampleSize / math.max(numItems,1L), 1.0)
        val candidates = ArrayBuffer.empty[(K, Float)]
        val imbalancedPartitions = mutable.Set.empty[Int]
        sketched.foreach { case (idx,n, sample) =>
          if (fraction * n > sampleSizePerPartition) {
            imbalancedPartitions += idx
          } else {
            // The weight is 1 over the sampling probability.
            val weight = (n.toDouble / sample.size).toFloat
            for (key<- sample) {
              candidates += ((key, weight))
            }
          }
        }

//针对不平衡的分区继续抽样
        if (imbalancedPartitions.nonEmpty) {
          // Re-sample imbalanced partitions with the desired sampling probability.
          val imbalanced =new PartitionPruningRDD(rdd.map(_._1), imbalancedPartitions.contains)
          val seed = byteswap32(-rdd.id– 1)
          val reSampled = imbalanced.sample(withReplacement =false, fraction, seed).collect()
          val weight = (1.0/ fraction).toFloat
          candidates ++= reSampled.map(x => (x, weight))
        }

//根据各个分区抽样的值来划分边界，其中weight值反应某个key的权重，权重越大，说明该key值越多
        RangePartitioner.determineBounds(candidates, partitions)
      }
    }
  }

其中sketch函数：

def sketch[K: ClassTag](
    rdd: RDD[K],
    sampleSizePerPartition: Int): (Long, Array[(Int, Int, Array[K])]) = {
  val shift = rdd.id
  // val classTagK = classTag[K] // to avoid serializing the entire partitioner object
  val sketched = rdd.mapPartitionsWithIndex { (idx, iter) =>
    val seed = byteswap32(idx ^ (shift <<16))
    //Reservoir:水塘抽样
    val (sample, n) = SamplingUtils.reservoirSampleAndCount(
      iter, sampleSizePerPartition, seed)
    Iterator((idx, n, sample))
  }.collect()
  val numItems = sketched.map(_._2.toLong).sum
  (numItems, sketched)
}

 

def reservoirSampleAndCount[T: ClassTag](
    input: Iterator[T],
    k: Int,
    seed: Long = Random.nextLong())
  : (Array[T], Int) = {
  val reservoir = newArray[T](k)
  // Put the first k elements in the reservoir.

  //先取前K个值

  vari = 0
  while (i < k && input.hasNext) {
    val item = input.next()
    reservoir(i) = item
    i += 1
  }

  // If we have consumed all the elements, return them. Otherwise do the replacement.
  if (i < k) {//如果没有取到，说明该分区少于K个值，则直接返回
    // If input size < k, trim the array to return only an array of input size.
    val trimReservoir =new Array[T](i)
    System.arraycopy(reservoir, 0, trimReservoir,0, i)
    (trimReservoir, i)
  } else {//否则按照水塘抽样遍历剩余的值
    // If input size > k, continue the sampling process.
    val rand = new XORShiftRandom(seed)
    while (input.hasNext) {
      val item = input.next()
      val replacementIndex = rand.nextInt(i)
      if (replacementIndex < k) {
        reservoir(replacementIndex) = item
      }
      i += 1
    }
    (reservoir, i)
  }
}

最后调用determineBounds来确定分界值：

def determineBounds[K: Ordering : ClassTag](
    candidates: ArrayBuffer[(K, Float)],
    partitions: Int): Array[K] = {
  val ordering = implicitly[Ordering[K]]
  val ordered = candidates.sortBy(_._1)
  val numCandidates = ordered.size
  val sumWeights = ordered.map(_._2.toDouble).sum
  val step = sumWeights / partitions
  var cumWeight = 0.0
  var target = step
  val bounds = ArrayBuffer.empty[K]
  var i = 0
  var j = 0
  var previousBound = Option.empty[K]
  while ((i < numCandidates) && (j < partitions –1)) {
    val (key, weight) = ordered(i)
    cumWeight += weight
    if (cumWeight > target) {//根据weight值来均衡划分分界
      // Skip duplicate values.
      if (previousBound.isEmpty || ordering.gt(key, previousBound.get)) {
        bounds += key
        target += step
        j += 1
        previousBound = Some(key)
      }
    }
    i += 1
  }
  bounds.toArray

}

其划分边界的具体流程如下：

Spark算子执行流程详解之一大数据

因此执行RDD的sortbykey操作，会导致对其key的至少一次扫描，比较耗时间，对外表现就是会执行一次action操作。

原创文章，作者：奋斗，如若转载，请注明出处：https://blog.ytso.com/9303.html

赞 (0)
0
上一篇 2021年7月19日
下一篇 2021年7月19日

相关推荐

发表回复

登录后才能评论