第2章 RDD编程（2.3）

2.3 TransFormation

基本RDD

Pair类型RDD

（伪集合操作交、并、补、笛卡尔积都支持）

2.3.1 map(func)

返回一个新的RDD，该RDD由每一个输入元素经过func函数转换后组成

package com.diyo.funtion

import org.apache.spark.rdd.RDD

import org.apache.spark.{SparkConf, SparkContext}

/**

  * map(func)

  * 返回一个新的RDD，该RDD由每一个输入元素经过func函数转换后组成

  */

object mapDemo extends App {

  /*map(func)

     返回一个新的RDD，该RDD由每一个输入元素经过func函数转换后组成

     scala> var source  = sc.parallelize(1 to 10)

     source: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[8] at parallelize at <console>:24

     scala> source.collect()

     res7: Array[Int] = Array(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

     scala> val mapadd = source.map(_ * 2)

     mapadd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[9] at map at <console>:26

     scala> mapadd.collect()

     res8: Array[Int] = Array(2, 4, 6, 8, 10, 12, 14, 16, 18, 20)

     ---------------------------------------------------------------------------------------------------

     map 是对 RDD 中的每个元素都执行一个指定的函数来产生一个新的 RDD

     任何 原 RDD 中的元素在新 RDD 中都有且只有一个元素与之对应。

     举例：

     scala>val a=sc.parallelize(1 to 9,3)

     scala>val b=a.map(x=>x*2)

     scala>a.collect res10:Array[Int]=Array(1,2,3,4,5,6,7,8,9)

     scala>b.collect res11:Array[Int]=Array(2,4,6,8,10,12,14,16,18)

     上述例子中把原 RDD 中每个元素都乘以 2 来产生一个新的 RDD。

      */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_map")

  val sc = new SparkContext(conf)

  val a: RDD[Int] = sc.parallelize(1 to 9 ,3)

  val b = a.map(x=>{

    println("map")

    x*2

  })

  a.foreach(println)

  println("a的分区数："+a.partitions.size)

  println(b.collect().mkString(""))

}

2.3.2 mapPartitions(func) 尽量使用mapPartitions

类似于map，但独立地在RDD的每一个分片上运行，因此在类型为T的RDD上运行时，func的函数类型必须是Iterator[T] => Iterator[U]。假设有N个元素，有M个分区，那么map的函数的将被调用N次,而mapPartitions被调用M次,一个函数一次处理所有分区

package com.diyo.funtion

/**

  * mapPartitions(func) 尽量使用mapPartitions

  * 类似于map，但独立地在RDD的每一个分片上运行，

  * 因此在类型为T的RDD上运行时，

  * func的函数类型必须是Iterator[T] => Iterator[U]

  * 假设有N个元素，有M个分区，那么map的函数的将被调用N次,

  * 而mapPartitions被调用M次,一个函数一次处理所有分区

  */

object mapPartitionsDemo extends App{

  /*

  mapPartitions(func)

    scala> val rdd = sc.parallelize( List( ("kpop","female") , ("zorro","male") , ("mobin","male") , ("lucy","female") ))

    rdd: org.apache.spark.rdd.RDD[(String, String)] = ParallelCollectionRDD[16] at parallelize at <console>:24

    scala> :paste

    // Entering paste mode (ctrl-D to finish)

    def partitionsFun(iter : Iterator[(String,String)]) : Iterator[String] = {

      var woman = List[String]()

      while (iter.hasNext){

        val next = iter.next()

       next match {

          case (_,"female") => woman = next._1 :: woman

         case _ =>

        }

     }

     woman.iterator

     }

    // Exiting paste mode, now interpreting.

    partitionsFun: (iter: Iterator[(String, String)])Iterator[String]

    scala> val result = rdd.mapPartitions(partitionsFun)

    result: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[17] at mapPartitions at <console>:28

      scala> result.collect()

      res13: Array[String] = Array(kpop, lucy)

   ------------------------------------------------------------

    mapPartitions 是 map 的一个变种。

    map 的输入函数是应用于 RDD 中每个元素，

    而 mapPartitions 的输入函数是应用于每个分区，

    也就是把每个分区中的内容作 为整体来处理的。

    它的函数定义为：

      def mapPartitions[U: ClassTag](f: Iterator[T] => Iterator[U], preservesPartitioning: Boolean=false):RDD[U]

      f 即为输入函数，它处理每个分区里面的内容。每个分区中的内容将以 Iterator[T] 传递给输入函数 f， f 的输出结果是 Iterator[U]

      最终的 RDD 由所有分区经过输入 函数处理后的结果合并起来的。

      举例：

      scala>val a=sc.parallelize(1 to 9,3)

      scala>def myfuncT:Iterator[(T,T)]={

              var res=List(T,T)

              var pre=iter.next

              while(iter.hasNext){

                val cur=iter.next

                res.::=(pre,cur)

                pre=cur }

              res.iterator }

      scala>a.mapPartitions(myfunc).collect

      res0:Array[(Int,Int)]=Array((2,3),(1,2),(5,6),(4,5),(8,9),(7,8))

      上述例子中的函数myfunc是把分区中一个元素和它的下一个元素组成一个Tuple。 因为分区中最后一个元素没有下一个元素了，所以(3,4)和(6,7)不在结果中。 mapPartitions 还有些变种，比如 mapPartitionsWithContext，它能把处理过程中的 一些状态信息传递给用户指定的输入函数。还有 mapPartitionsWithIndex，它能 把分区的 index传递给用户指定的输入函数。

      */

}

2.3.3 glom

将每一个分区形成一个数组，形成新的RDD类型时RDD[Array[T]]

scala> val rdd = sc.parallelize(1 to 16,4)

rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[65] at parallelize at <console>:24

scala> rdd.glom().collect()

res25: Array[Array[Int]] = Array(Array(1, 2, 3, 4), Array(5, 6, 7, 8), Array(9, 10, 11, 12), Array(13, 14, 15, 16))

2.3.4 flatMap(func) map后再扁平化

类似于map，但是每一个输入元素可以被映射为0或多个输出元素（所以func应该返回一个序列，而不是单一元素）

package com.diyo.funtion

import org.apache.spark.{SparkConf, SparkContext}

/**

  * flatMap(func) map后再扁平化

  * 类似于map，但是每一个输入元素可以被映射为0或多个输出元素（所以func应该返回一个序列，而不是单一元素）

  */

object flatMapDemo extends App {

  /*

  scala> val sourceFlat = sc.parallelize(1 to 5)

  sourceFlat: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[12] at parallelize at <console>:24

    scala> sourceFlat.collect()

    res11: Array[Int] = Array(1, 2, 3, 4, 5)

    scala> val flatMap = sourceFlat.flatMap(1 to _)

    flatMap: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[13] at flatMap at <console>:26

      scala> flatMap.collect()

      res12: Array[Int] = Array(1, 1, 2, 1, 2, 3, 1, 2, 3, 4, 1, 2, 3, 4, 5)

  */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_flatMap")

  val sc = new SparkContext(conf)

  val sourceFlat = sc.parallelize(1 to 5)

  val flatMap = sourceFlat.flatMap(1 to _)      //(x => (1 to x))

  println(flatMap.collect().mkString(""))

}

2.3.5 filter(func)

返回一个新的RDD，该RDD由经过func函数计算后返回值为true的输入元素组成

package com.diyo.funtion

import org.apache.spark.{SparkConf, SparkContext}

/**

  * filter(func)

  * 返回一个新的RDD，该RDD由经过func函数计算后返回值为true的输入元素组成

  */

object filterDemo extends App {

  /*返回一个新的RDD，该RDD由经过func函数计算后返回值为true的输入元素组成

  scala> var sourceFilter = sc.parallelize(Array("xiaoming","xiaojiang","xiaohe","dazhi"))

  sourceFilter: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[10] at parallelize at <console>:24

  scala> val filter = sourceFilter.filter(_.contains("xiao"))

  filter: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[11] at filter at <console>:26

  scala> sourceFilter.collect()

  res9: Array[String] = Array(xiaoming, xiaojiang, xiaohe, dazhi)

  scala> filter.collect()

  res10: Array[String] = Array(xiaoming, xiaojiang, xiaohe)

  --------------------------------------------------------

  filter 是对 RDD 中的每个元素都执行一个指定的函数来过滤产生一个新的 RDD。

  任何原 RDD 中的元素在新 RDD 中都有且只有一个元素与之对应。

  val rdd=sc.parallelize(List(1,2,3,4,5,6))

  val filterRdd=rdd.filter(_>5)

  filterRdd.collect()//返回所有大于 5 的数据的一个 Array， Array(6,8,10,12)

  */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_flatMap")

  val sc = new SparkContext(conf)

  val rdd = sc.parallelize(List(1, 2, 3, 4, 5, 6))

  val filterRdd = rdd.filter(x => x > 5)

  println(filterRdd.collect().mkString(""))

2.3.6 mapPartitionsWithIndex(func)

类似于mapPartitions，但func带有一个整数参数表示分片的索引值，因此在类型为T的RDD上运行时，func的函数类型必须是(Int, Interator[T]) => Iterator[U]

package com.diyo.funtion

import org.apache.spark.{SparkConf, SparkContext}

/**

  * mapPartitionsWithIndex(func)

  * 类似于mapPartitions，但func带有一个整数参数表示分片的索引值，

  * 因此在类型为T的RDD上运行时，func的函数类型必须是(Int, Interator[T]) => Iterator[U]

  */

object mapPartitionsWithIndexDemo extends App {

/*

  类似于mapPartitions，但func带有一个整数参数表示分片的索引值，因此在类型为T的RDD上运行时，func的函数类型必须是(Int, Interator[T]) => Iterator[U]

  scala> val rdd = sc.parallelize(List(("kpop","female"),("zorro","male"),("mobin","male"),("lucy","female")))

  rdd: org.apache.spark.rdd.RDD[(String, String)] = ParallelCollectionRDD[18] at parallelize at <console>:24

    scala> :paste

    // Entering paste mode (ctrl-D to finish)

    def partitionsFun(index : Int, iter : Iterator[(String,String)]) : Iterator[String] = {

    var woman = List[String]()

    while (iter.hasNext){

      val next = iter.next()

      next match {

        case (_,"female") => woman = "["+index+"]"+next._1 :: woman

        case _ =>

      }

    }

    woman.iterator

    }

    // Exiting paste mode, now interpreting.

    partitionsFun: (index: Int, iter: Iterator[(String, String)])Iterator[String]

    scala> val result = rdd.mapPartitionsWithIndex(partitionsFun)

    result: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[19] at mapPartitionsWithIndex at <console>:28

      scala> result.collect()

      res14: Array[String] = Array([0]kpop, [3]lucy)

      ----------------------------------------------

   def mapPartitionsWithIndex[U](f: (Int, Iterator[T]) => Iterator[U], preservesPartitioning:Boolean=false)(implicitarg0:ClassTag[U]):RDD[U]

   函数作用同 mapPartitions，不过提供了两个参数，第一个参数为分区的索引。

    var rdd1 = sc.makeRDD(1 to 5,2) //rdd1 有两个分区

    var rdd2=rdd1.mapPartitionsWithIndex{

      (x,iter)=>{

       var result = ListString

       var i=0

       while(iter.hasNext){

        i+=iter.next() }

      result.::(x+"|"+i).iterator

      }

    } //rdd2 将 rdd1 中每个分区的数字累加，并在每个分区的累加结果前面加了分区 索引

    scala>rdd2.collect

    res13:Array[String]=Array(0|3,1|12)

  */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_mapPartitionsWithIndex")

  val sc = new SparkContext(conf)

  val rdd = sc.parallelize(Array(1,2,3,4,5,6),2)

  val a = rdd.mapPartitionsWithIndex((x,y) => Iterator(x+":"+y.mkString("")))   //(x,y) x为分区号,y为分区中内容

  println(a.collect().mkString(","))      //0:123,1:456

}

2.3.7 sample(withReplacement, fraction, seed)

以指定的随机种子随机抽样出数量为fraction的数据，withReplacement表示是抽出的数据是否放回，true为有放回的抽样，false为无放回的抽样，seed用于指定随机数生成器种子。例子从RDD中随机且有放回的抽出50%的数据，随机种子值为3（即可能以1 2 3的其中一个起始值）

package com.diyo.funtion

import org.apache.spark.{SparkConf, SparkContext}

/**

  * sample(withReplacement, fraction, seed)

  * withReplacement表示是抽出的数据是否放回，true为有放回的抽样，false为无放回的抽样，

  * 以指定的随机种子随机抽样出数量为fraction的数据，

  * seed用于指定随机数生成器种子。

  * 例子从RDD中随机且有放回的抽出50%的数据，随机种子值为3（即可能以1 2 3的其中一个起始值）

  * sample算子时用来抽样用的，其有3个参数

  *

  * withReplacement：表示抽出样本后是否在放回去，true表示会放回去，这也就意味着抽出的样本可能有重复

  *

  * fraction ：抽出多少，这是一个double类型的参数,0-1之间，eg:0.3表示抽出30%

  *

  * seed：表示一个种子，根据这个seed随机抽取，一般情况下只用前两个参数就可以，

  * 那么这个参数是干嘛的呢，这个参数一般用于调试，有时候不知道是程序出问题还是数据出了问题，就可以将这个参数设置为定值

  */

object sampleDemo extends App {

  /*

  scala> val rdd = sc.parallelize(1 to 10)

  rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[20] at parallelize at <console>:24

    scala> rdd.collect()

    res15: Array[Int] = Array(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

    scala> var sample1 = rdd.sample(true,0.4,2)

    sample1: org.apache.spark.rdd.RDD[Int] = PartitionwiseSampledRDD[21] at sample at <console>:26

    scala> sample1.collect()

    res16: Array[Int] = Array(1, 2, 2, 7, 7, 8, 9)

    scala> var sample2 = rdd.sample(false,0.2,3)

    sample2: org.apache.spark.rdd.RDD[Int] = PartitionwiseSampledRDD[22] at sample at <console>:26

    scala> sample2.collect()

    res17: Array[Int] = Array(1, 9)

  */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_sample")

  val sc = new SparkContext(conf)

  val rdd = sc.parallelize(1 to 10)

  val a = rdd.sample(true,0.3)

  println(a.collect().mkString(""))

}

2.3.8 distinct([numTasks]))

对源RDD进行去重后返回一个新的RDD. 默认情况下，只有8个并行任务来操作，但是可以传入一个可选的numTasks参数改变它。

package com.diyo.funtion

import org.apache.spark.{SparkConf, SparkContext}

/**

  * distinct([numTasks]))

  * 对源RDD进行去重后返回一个新的RDD. 默认情况下，只有8个并行任务来操作，但是可以传入一个可选的numTasks参数改变它。

  */

object distinctDemo extends App {

  /*

  scala> val distinctRdd = sc.parallelize(List(1,2,1,5,2,9,6,1))

  distinctRdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[34] at parallelize at <console>:24

  scala> val unionRDD = distinctRdd.distinct()

  unionRDD: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[37] at distinct at <console>:26

  scala> unionRDD.collect()

  [Stage 16:> (0 + 4) [Stage 16:=============================>                            (2 + 2)                                                                             res20: Array[Int] = Array(1, 9, 5, 6, 2)

  scala> val unionRDD = distinctRdd.distinct(2)

  unionRDD: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[40] at distinct at <console>:26

  scala> unionRDD.collect()

  res21: Array[Int] = Array(6, 2, 1, 9, 5)

  --------------------------------------------------------

  distinct

  去重

  val rdd1 = sc.parallelize(List(5,6,4,3))

  val rdd2 = sc.parallelize(List(1,2,3,4)) //求并集

  val rdd3 = rdd1.union(rdd2) //去重输出

  rdd3.distinct.collect

  */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_distinct")

  val sc = new SparkContext(conf)

  val rdd = sc.parallelize(Array(1,2,2,1,3,4,5,5,5,6))

  val a = rdd.distinct()

//  val a = rdd.distinct(2)     //参数为Task数

  println(a.collect().mkString(""))

}

2.3.9 partitionBy

对RDD进行分区操作，如果原有的partionRDD和现有的partionRDD是一致的话就不进行分区，否则会生成ShuffleRDD。

package com.diyo.funtion

import org.apache.spark.{HashPartitioner, SparkConf, SparkContext}

/**

  * partitionBy

  * 对RDD进行分区操作，如果原有的partionRDD和现有的partionRDD是一致的话就不进行分区， 否则会生成ShuffleRDD.

  */

object partitionByDemo extends App {

  /*

  scala> val rdd = sc.parallelize(Array((1,"aaa"),(2,"bbb"),(3,"ccc"),(4,"ddd")),4)

  rdd: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[44] at parallelize at <console>:24

  scala> rdd.partitions.size

  res24: Int = 4

  scala> var rdd2 = rdd.partitionBy(new org.apache.spark.HashPartitioner(2))

  rdd2: org.apache.spark.rdd.RDD[(Int, String)] = ShuffledRDD[45] at partitionBy at <console>:26

  scala> rdd2.partitions.size

  res25: Int = 2

  */

  val conf = new SparkConf().setMaster("local[*]").setAppName("dzy_partitionBy")

  val sc = new SparkContext(conf)

  val rdd = sc.parallelize(Array((1,"aaa"),(2,"bbb"),(3,"ccc"),(4,"ddd")))

  println(rdd.collect().mkString(""))

  val rdd2 = rdd.partitionBy(new HashPartitioner(2))

  println(rdd2.collect().mkString(""))

}

2.3.10 coalesce(numPartitions)

与repartition的区别: repartition(numPartitions:Int):RDD[T]和coalesce(numPartitions:Int，shuffle:Boolean=false):RDD[T] repartition只是coalesce接口中shuffle为true的实现.

缩减分区数，用于大数据集过滤后，提高小数据集的执行效率。

scala> val rdd = sc.parallelize(1 to 16,4)

rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[54] at parallelize at <console>:24

scala> rdd.partitions.size

res20: Int = 4

scala> val coalesceRDD = rdd.coalesce(3)

coalesceRDD: org.apache.spark.rdd.RDD[Int] = CoalescedRDD[55] at coalesce at <console>:26

scala> coalesceRDD.partitions.size

res21: Int = 3

2.3.11 repartition(numPartitions)

根据分区数，从新通过网络随机洗牌所有数据。

scala> val rdd = sc.parallelize(1 to 16,4)

rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[56] at parallelize at <console>:24

scala> rdd.partitions.size

res22: Int = 4

scala> val rerdd = rdd.repartition(2)

rerdd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[60] at repartition at <console>:26

scala> rerdd.partitions.size

res23: Int = 2

scala> val rerdd = rdd.repartition(4)

rerdd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[64] at repartition at <console>:26

scala> rerdd.partitions.size

res24: Int = 4

2.3.12 repartitionAndSortWithinPartitions(partitioner)

repartitionAndSortWithinPartitions函数是repartition函数的变种，与repartition函数不同的是，repartitionAndSortWithinPartitions在给定的partitioner内部进行排序，性能比repartition要高。

2.3.13 sortBy(func,[ascending], [numTasks])

用func先对数据进行处理，按照处理后的数据比较结果排序。

scala> val rdd = sc.parallelize(List(1,2,3,4))

rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[21] at parallelize at <console>:24

scala> rdd.sortBy(x => x).collect()

res11: Array[Int] = Array(1, 2, 3, 4)

scala> rdd.sortBy(x => x%3).collect()

res12: Array[Int] = Array(3, 4, 1, 2)

2.3.14 union(otherDataset)

对源RDD和参数RDD求并集后返回一个新的RDD 不去重

scala> val rdd1 = sc.parallelize(1 to 5)

rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[23] at parallelize at <console>:24

scala> val rdd2 = sc.parallelize(5 to 10)

rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[24] at parallelize at <console>:24

scala> val rdd3 = rdd1.union(rdd2)

rdd3: org.apache.spark.rdd.RDD[Int] = UnionRDD[25] at union at <console>:28

scala> rdd3.collect()

res18: Array[Int] = Array(1, 2, 3, 4, 5, 5, 6, 7, 8, 9, 10)

2.3.15 subtract (otherDataset)

计算差的一种函数，去除两个RDD中相同的元素，不同的RDD将保留下来

scala> val rdd = sc.parallelize(3 to 8)

rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[70] at parallelize at <console>:24

scala> val rdd1 = sc.parallelize(1 to 5)

rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[71] at parallelize at <console>:24

scala> rdd.subtract(rdd1).collect()

res27: Array[Int] = Array(8, 6, 7)

2.3.16 intersection(otherDataset)

对源RDD和参数RDD求交集后返回一个新的RDD

scala> val rdd1 = sc.parallelize(1 to 7)

rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[26] at parallelize at <console>:24

scala> val rdd2 = sc.parallelize(5 to 10)

rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[27] at parallelize at <console>:24

scala> val rdd3 = rdd1.intersection(rdd2)

rdd3: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[33] at intersection at <console>:28

scala> rdd3.collect()

res19: Array[Int] = Array(5, 6, 7)

2.3.17 cartesian(otherDataset)

笛卡尔积

scala> val rdd1 = sc.parallelize(1 to 3)

rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[47] at parallelize at <console>:24

scala> val rdd2 = sc.parallelize(2 to 5)

rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[48] at parallelize at <console>:24

scala> rdd1.cartesian(rdd2).collect()

res17: Array[(Int, Int)] = Array((1,2), (1,3), (1,4), (1,5), (2,2), (2,3), (2,4), (2,5), (3,2), (3,3), (3,4), (3,5))

2.3.18 pipe(command, [envVars])

管道，对于每个分区，都执行一个perl或者shell脚本，返回输出的RDD

Shell脚本

#!/bin/sh

echo "AA"

while read LINE; do

echo ">>>"${LINE}

done

scala> val rdd = sc.parallelize(List("hi","Hello","how","are","you"),1)

rdd: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[50] at parallelize at <console>:24

scala> rdd.pipe("/home/bigdata/pipe.sh").collect()

res18: Array[String] = Array(AA, >>>hi, >>>Hello, >>>how, >>>are, >>>you)

scala> val rdd = sc.parallelize(List("hi","Hello","how","are","you"),2)

rdd: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[52] at parallelize at <console>:24

scala> rdd.pipe("/home/bigdata/pipe.sh").collect()

res19: Array[String] = Array(AA, >>>hi, >>>Hello, AA, >>>how, >>>are, >>>you)

pipe.sh:

#!/bin/sh

echo "AA"

while read LINE; do

echo ">>>"${LINE}

done

2.3.19 join(otherDataset, [numTasks])

在类型为(K,V)和(K,W)的RDD上调用，返回一个相同key对应的所有元素对在一起的(K,(V,W))的RDD

scala> val rdd = sc.parallelize(Array((1,"a"),(2,"b"),(3,"c")))

rdd: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[32] at parallelize at <console>:24

scala> val rdd1 = sc.parallelize(Array((1,4),(2,5),(3,6)))

rdd1: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[33] at parallelize at <console>:24

scala> rdd.join(rdd1).collect()

res13: Array[(Int, (String, Int))] = Array((1,(a,4)), (2,(b,5)), (3,(c,6)))

2.3.20 cogroup(otherDataset, [numTasks])

在类型为(K,V)和(K,W)的RDD上调用，返回一个(K,(Iterable<V>,Iterable<W>))类型的RDD

scala> val rdd = sc.parallelize(Array((1,"a"),(2,"b"),(3,"c")))

rdd: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[37] at parallelize at <console>:24

scala> val rdd1 = sc.parallelize(Array((1,4),(2,5),(3,6)))

rdd1: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[38] at parallelize at <console>:24

scala> rdd.cogroup(rdd1).collect()

res14: Array[(Int, (Iterable[String], Iterable[Int]))] = Array((1,(CompactBuffer(a),CompactBuffer(4))), (2,(CompactBuffer(b),CompactBuffer(5))), (3,(CompactBuffer(c),CompactBuffer(6))))

scala> val rdd2 = sc.parallelize(Array((4,4),(2,5),(3,6)))

rdd2: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[41] at parallelize at <console>:24

scala> rdd.cogroup(rdd2).collect()

res15: Array[(Int, (Iterable[String], Iterable[Int]))] = Array((4,(CompactBuffer(),CompactBuffer(4))), (1,(CompactBuffer(a),CompactBuffer())), (2,(CompactBuffer(b),CompactBuffer(5))), (3,(CompactBuffer(c),CompactBuffer(6))))

scala> val rdd3 = sc.parallelize(Array((1,"a"),(1,"d"),(2,"b"),(3,"c")))

rdd3: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[44] at parallelize at <console>:24

scala> rdd3.cogroup(rdd2).collect()

[Stage 36:> (0 + 0) res16: Array[(Int, (Iterable[String], Iterable[Int]))] = Array((4,(CompactBuffer(),CompactBuffer(4))), (1,(CompactBuffer(d, a),CompactBuffer())), (2,(CompactBuffer(b),CompactBuffer(5))), (3,(CompactBuffer(c),CompactBuffer(6))))

2.3.21 reduceByKey(func, [numTasks])

在一个(K,V)的RDD上调用，返回一个(K,V)的RDD，使用指定的reduce函数，将相同key的值聚合到一起，reduce任务的个数可以通过第二个可选的参数来设置。

scala> val rdd = sc.parallelize(List(("female",1),("male",5),("female",5),("male",2)))

rdd: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[46] at parallelize at <console>:24

scala> val reduce = rdd.reduceByKey((x,y) => x+y)

reduce: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[47] at reduceByKey at <console>:26

scala> reduce.collect()

res29: Array[(String, Int)] = Array((female,6), (male,7))

2.3.22 groupByKey

groupByKey也是对每个key进行操作，但只生成一个sequence。

scala> val words = Array("one", "two", "two", "three", "three", "three")

words: Array[String] = Array(one, two, two, three, three, three)

scala> val wordPairsRDD = sc.parallelize(words).map(word => (word, 1))

wordPairsRDD: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[4] at map at <console>:26

scala> val group = wordPairsRDD.groupByKey()

group: org.apache.spark.rdd.RDD[(String, Iterable[Int])] = ShuffledRDD[5] at groupByKey at <console>:28

scala> group.collect()

res1: Array[(String, Iterable[Int])] = Array((two,CompactBuffer(1, 1)), (one,CompactBuffer(1)), (three,CompactBuffer(1, 1, 1)))

scala> group.map(t => (t._1, t._2.sum))

res2: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[6] at map at <console>:31

scala> res2.collect()

res3: Array[(String, Int)] = Array((two,2), (one,1), (three,3))

scala> val map = group.map(t => (t._1, t._2.sum))

map: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[7] at map at <console>:30

scala> map.collect()

res4: Array[(String, Int)] = Array((two,2), (one,1), (three,3))

2.3.23 combineByKey[C]

( createCombiner: V => C, mergeValue: (C, V) => C, mergeCombiners: (C, C) => C)

对相同K，把V合并成一个集合。

createCombiner: combineByKey() 会遍历分区中的所有元素，因此每个元素的键要么还没有遇到过，要么就和之前的某个元素的键相同。如果这是一个新的元素,combineByKey() 会使用一个叫作 createCombiner() 的函数来创建
那个键对应的累加器的初始值

mergeValue: 如果这是一个在处理当前分区之前已经遇到的键，它会使用 mergeValue() 方法将该键的累加器对应的当前值与这个新的值进行合并

mergeCombiners: 由于每个分区都是独立处理的，因此对于同一个键可以有多个累加器。如果有两个或者更多的分区都有对应同一个键的累加器，就需要使用用户提供的 mergeCombiners() 方法将各个分区的结果进行合并。

scala> val scores = Array(("Fred", 88), ("Fred", 95), ("Fred", 91), ("Wilma", 93), ("Wilma", 95), ("Wilma", 98))

scores: Array[(String, Int)] = Array((Fred,88), (Fred,95), (Fred,91), (Wilma,93), (Wilma,95), (Wilma,98))

scala> val input = sc.parallelize(scores)

input: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[52] at parallelize at <console>:26

scala> val combine = input.combineByKey(

| (v)=>(v,1),

| (acc:(Int,Int),v)=>(acc._1+v,acc._2+1),

| (acc1:(Int,Int),acc2:(Int,Int))=>(acc1._1+acc2._1,acc1._2+acc2._2))

combine: org.apache.spark.rdd.RDD[(String, (Int, Int))] = ShuffledRDD[53] at combineByKey at <console>:28

scala> val result = combine.map{

| case (key,value) => (key,value._1/value._2.toDouble)}

result: org.apache.spark.rdd.RDD[(String, Double)] = MapPartitionsRDD[54] at map at <console>:30

scala> result.collect()

res33: Array[(String, Double)] = Array((Wilma,95.33333333333333), (Fred,91.33333333333333))

2.3.24 aggregateByKey

(zeroValue:U,[partitioner: Partitioner]) (seqOp: (U, V) => U,combOp: (U, U) => U)

在kv对的RDD中，，按key将value进行分组合并，合并时，将每个value和初始值作为seq函数的参数，进行计算，返回的结果作为一个新的kv对，然后再将结果按照key进行合并，最后将每个分组的value传递给combine函数进行计算（先将前两个value进行计算，将返回结果和下一个value传给combine函数，以此类推），将key与计算结果作为一个新的kv对输出。

seqOp函数用于在每一个分区中用初始值逐步迭代value，combOp函数用于合并每个分区中的结果。

scala> val rdd = sc.parallelize(List((1,3),(1,2),(1,4),(2,3),(3,6),(3,8)),3)

rdd: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[12] at parallelize at <console>:24

scala> val agg = rdd.aggregateByKey(0)(math.max(_,_),_+_)

agg: org.apache.spark.rdd.RDD[(Int, Int)] = ShuffledRDD[13] at aggregateByKey at <console>:26

scala> agg.collect()

res7: Array[(Int, Int)] = Array((3,8), (1,7), (2,3))

scala> agg.partitions.size

res8: Int = 3

scala> val rdd = sc.parallelize(List((1,3),(1,2),(1,4),(2,3),(3,6),(3,8)),1)

rdd: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[10] at parallelize at <console>:24

scala> val agg = rdd.aggregateByKey(0)(math.max(_,_),_+_).collect()

agg: Array[(Int, Int)] = Array((1,4), (3,8), (2,3))

2.3.25 foldByKey

(zeroValue: V)(func: (V, V) => V): RDD[(K, V)]

aggregateByKey的简化操作，seqop和combop相同

scala> val rdd = sc.parallelize(List((1,3),(1,2),(1,4),(2,3),(3,6),(3,8)),3)

rdd: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[91] at parallelize at <console>:24

scala> val agg = rdd.foldByKey(0)(_+_)

agg: org.apache.spark.rdd.RDD[(Int, Int)] = ShuffledRDD[92] at foldByKey at <console>:26

scala> agg.collect()

res61: Array[(Int, Int)] = Array((3,14), (1,9), (2,3))

2.3.26 sortByKey([ascending], [numTasks])

在一个(K,V)的RDD上调用，K必须实现Ordered接口，返回一个按照key进行排序的(K,V)的RDD

scala> val rdd = sc.parallelize(Array((3,"aa"),(6,"cc"),(2,"bb"),(1,"dd")))

rdd: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[14] at parallelize at <console>:24

scala> rdd.sortByKey(true).collect()

res9: Array[(Int, String)] = Array((1,dd), (2,bb), (3,aa), (6,cc))

scala> rdd.sortByKey(false).collect()

res10: Array[(Int, String)] = Array((6,cc), (3,aa), (2,bb), (1,dd))

2.3.27 mapValues

针对于(K,V)形式的类型只对V进行操作

scala> val rdd3 = sc.parallelize(Array((1,"a"),(1,"d"),(2,"b"),(3,"c")))

rdd3: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[67] at parallelize at <console>:24

scala> rdd3.mapValues(_+"|||").collect()

res26: Array[(Int, String)] = Array((1,a|||), (1,d|||), (2,b|||), (3,c|||))