Spark算子执行流程详解之三

10.aggregate

用与聚合RDD中的元素，先使用seqOp将RDD中每个分区中的T类型元素聚合成U类型，再使用combOp将之前每个分区聚合后的U类型聚合成U类型，特别注意seqOp和combOp都会使用zeroValue的值，zeroValue的类型为U，

def aggregate[U: ClassTag](zeroValue:U)(seqOp: (U,T) =>U, combOp: (U,U) =>U): U = withScope {
  // Clone the zero value since we will also be serializing it as part of tasks
  var jobResult = Utils.clone(zeroValue, sc.env.serializer.newInstance())
  val cleanSeqOp = sc.clean(seqOp)
  val cleanCombOp = sc.clean(combOp)

// zeroValue即初始值，aggregatePartition是在excutor上执行的

val aggregatePartition = (it:Iterator[T]) => it.aggregate(zeroValue)(cleanSeqOp, cleanCombOp)

// jobResult即初始值，其合并每个分区的结果是在driver端执行的
  val mergeResult = (index: Int, taskResult:U) => jobResult = combOp(jobResult, taskResult)

  sc.runJob(this, aggregatePartition, mergeResult)
  jobResult
}

例如：

var rdd1 = sc.makeRDD(1 to 10,2)

##第一个分区中包含5,4,3,2,1

##第二个分区中包含10,9,8,7,6

scala> rdd1.aggregate(1)(

     |           {(x : Int,y : Int) => x + y},

     |           {(a : Int,b : Int) => a + b}

     |     )

res17: Int = 58

为什么是58呢？且看下面的执行流程：

11.fold

简化的aggregate，将aggregate中的seqOp和combOp使用同一个函数op。

/**
 * Aggregate the elements of each partition, and then the results for all the partitions, using a
 * given associative and commutative function and a neutral "zero value". The function
 * op(t1, t2) is allowed to modify t1 and return it as its result value to avoid object
 * allocation; however, it should not modify t2.
 *
 * This behaves somewhat differently from fold operations implemented for non-distributed
 * collections in functional languages like Scala. This fold operation may be applied to
 * partitions individually, and then fold those results into the final result, rather than
 * apply the fold to each element sequentially in some defined ordering. For functions
 * that are not commutative, the result may differ from that of a fold applied to a
 * non-distributed collection.
 */
def fold(zeroValue: T)(op: (T, T) => T): T= withScope {
  // Clone the zero value since we will also be serializing it as part of tasks
  var jobResult = Utils.clone(zeroValue, sc.env.closureSerializer.newInstance())
  val cleanOp = sc.clean(op)

//先在excutor上针对分区进行一次fold操作
  val foldPartition = (iter: Iterator[T]) => iter.fold(zeroValue)(cleanOp)

//然后在driver端合并每个分区上的结果
  val mergeResult = (index: Int, taskResult:T) => jobResult = op(jobResult, taskResult)
  sc.runJob(this, foldPartition, mergeResult)
  jobResult

}

例如可以将aggregate小节里面的例子操作转化为fold操作：

scala>var rdd1 = sc.makeRDD(1 to 10,2)

##第一个分区中包含5,4,3,2,1

##第二个分区中包含10,9,8,7,6

scala> rdd1.fold(1)(

     |       (x,y) => x + y   

     |     )

res19: Int = 58

##结果同上面使用aggregate的第一个例子一样，即：

scala> rdd1.aggregate(1)(

     |           {(x,y) => x + y},

     |           {(a,b) => a + b}

     |     )

res20: Int = 58

 

12.treeAggregate

分层进行aggregate，由于aggregate的时候其分区的结算结果是传输到driver端再进行合并的，如果分区比较多，计算结果返回的数据量比较大的话，那么driver端需要缓存大量的中间结果，这样就会加大driver端的计算能力，因此treeAggregate把分区计算结果的合并仍旧放在excutor端进行，将结果在excutor端不断合并缩小返回driver的数据量，最后再driver端进行最后一次合并。

/**
 * Aggregates the elements of this RDD in a multi-level tree pattern.
 *
 * @param depth suggested depth of the tree (default: 2)
 * @see [[org.apache.spark.rdd.RDD#aggregate]]
 */
def treeAggregate[U: ClassTag](zeroValue:U)(
    seqOp: (U, T) =>U,
    combOp: (U, U) =>U,
    depth: Int = 2): U = withScope {
  require(depth >= 1, s"Depth must be greater than or equal to 1 but got$depth.")
  if (partitions.length == 0) {
    Utils.clone(zeroValue, context.env.closureSerializer.newInstance())
  } else {
    val cleanSeqOp = context.clean(seqOp)
    val cleanCombOp = context.clean(combOp)

//针对初始分区的聚合函数
    val aggregatePartition =
      (it: Iterator[T]) => it.aggregate(zeroValue)(cleanSeqOp, cleanCombOp)

//针对初始的各分区先进行部分聚合
    var partiallyAggregated = mapPartitions(it =>Iterator(aggregatePartition(it)))
    var numPartitions = partiallyAggregated.partitions.length

//根据传入的depth计算出需要迭代计算的程度
    val scale = math.max(math.ceil(math.pow(numPartitions,1.0/ depth)).toInt, 2)
    // If creating an extra level doesn't help reduce
    // the wall-clock time, we stop tree aggregation.
    while (numPartitions > scale + numPartitions / scale) {//计算迭代的程度
      numPartitions /= scale
      val curNumPartitions = numPartitions

//减少分区个数，合并部分分区的结果
      partiallyAggregated = partiallyAggregated.mapPartitionsWithIndex {
        (i, iter) => iter.map((i % curNumPartitions, _))
      }.reduceByKey(new HashPartitioner(curNumPartitions), cleanCombOp).values
    }

//执行最后一次reduce，返回最终结果
    partiallyAggregated.reduce(cleanCombOp)
  }
}

例如：

scala> def seq(a:Int,b:Int):Int={

             | a+b}

        seq: (a: Int, b: Int)Int

        scala> def comb(a:Int,b:Int):Int={

             | a+b}

        comb: (a: Int, b: Int)Int

        val z =sc.parallelize(List(1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18),9)

        scala> z.treeAggregate(0)(seq,comb,2)

        res1: Int = 171

其具体的执行过程如下：

13.reduce

RDD中元素前两个传给输入函数，产生一个新的return值，新产生的return值与RDD中下一个元素（第三个元素）组成两个元素，再被传给输入函数，直到最后只有一个值为止。

/**
 * Reduces the elements of this RDD using the specified commutative and
 * associative binary operator.
 */
def reduce(f: (T,T) =>T): T = withScope {
  val cleanF = sc.clean(f)

//定义一个遍历partition的函数，这是在excutor端执行的
  val reducePartition: Iterator[T] => Option[T] = iter => {
    if (iter.hasNext) {

//reduceLeft从左往后遍历
      Some(iter.reduceLeft(cleanF))
    } else {
      None
    }
  }
  var jobResult: Option[T] = None

//定义一个driver端处理分区计算结果的函数，这是在driver端执行的
  val mergeResult = (index: Int, taskResult: Option[T]) => {
    if (taskResult.isDefined) {
      jobResult = jobResult match {
        case Some(value) =>Some(f(value, taskResult.get))
        case None => taskResult
      }
    }
  }
  sc.runJob(this, reducePartition, mergeResult)
  // Get the final result out of our Option, or throw an exception if the RDD was empty

//将结果返回
  jobResult.getOrElse(throw new UnsupportedOperationException("empty collection"))
}

例如：

val c = sc.parallelize(1 to 10 , 2)

c.reduce((x, y) => x + y)//结果55

具体执行流程如下：

14.max

返回最大值，其排序方法对象默认的排序方法

/**
 * Returns the max of this RDD as defined by the implicit Ordering[T].
 * @return the maximum element of the RDD
 * */
def max()(implicitord:Ordering[T]):T= withScope {
  this.reduce(ord.max)

}

其本质就是定义个排序的方法，然后调用reduce操作，实例如下：

scala>var rdd1 = sc.makeRDD(1 to 10,2)

##第一个分区中包含5,4,3,2,1

##第二个分区中包含10,9,8,7,6

scala> rdd1.max()

res19: Int = 10

其执行流程如下：

15.min

返回最小值，其排序方法对象默认的排序方法

/**
 * Returns the min of this RDD as defined by the implicit Ordering[T].
 * @return the maximum element of the RDD
 * */
def min()(implicitord:Ordering[T]):T= withScope {
  this.reduce(ord.min)

}

其本质就是定义个排序的方法，然后调用reduce操作，实例如下：

scala>var rdd1 = sc.makeRDD(1 to 10,2)

##第一个分区中包含5,4,3,2,1

##第二个分区中包含10,9,8,7,6

scala> rdd1.min()

res19: Int = 1

其执行流程如下：

16.treeReduce

类似于treeAggregate，利用在excutor端进行多次aggregate来缩小driver的计算开销

/**
 * Reduces the elements of this RDD in a multi-level tree pattern.
 *
 * @param depth suggested depth of the tree (default: 2)
 * @see [[org.apache.spark.rdd.RDD#reduce]]
 */
def treeReduce(f: (T,T) =>T, depth: Int =2):T = withScope {
  require(depth >= 1, s"Depth must be greater than or equal to 1 but got$depth.")
  val cleanF = context.clean(f)

//针对初始分区的reduce函数
  val reducePartition: Iterator[T] => Option[T] = iter => {
    if (iter.hasNext) {
      Some(iter.reduceLeft(cleanF))
    } else {
      None
    }
  }

//针对初始的各分区先进行部分reduce
  val partiallyReduced = mapPartitions(it =>Iterator(reducePartition(it)))
  val op: (Option[T], Option[T]) => Option[T] = (c, x) => {
    if (c.isDefined && x.isDefined) {
      Some(cleanF(c.get, x.get))
    } else if (c.isDefined) {
      c
    } else if (x.isDefined) {
      x
    } else {
      None
    }
  }

//最终调用的还是treeAggregate方法
  partiallyReduced.treeAggregate(Option.empty[T])(op, op, depth)
    .getOrElse(throw new UnsupportedOperationException("empty collection"))

}

treeReduce函数先是针对每个分区利用scala的reduceLeft函数进行计算；最后，在将局部合并的RDD进行treeAggregate计算，这里的seqOp和combOp一样，初值为空。在实际应用中，可以用treeReduce来代替reduce，主要是用于单个reduce操作开销比较大，而treeReduce可以通过调整深度来控制每次reduce的规模。其具体的执行流程不再详细叙述，可以参考treeAggregate方法。