04、常用RDD操作整理

常用Transformation

注:某些函数只有PairRDD只有,而普通的RDD则没有,比如gropuByKeyreduceByKeysortByKeyjoincogroup等函数要根据Key进行分组或直接操作

RDD基本转换:

RDD[U] map(f: T => U)

T:原RDD中元素类型

U:新RDD中元素类型

函数将T元素转换为新的U元素

rdd.map(x => x + 1)

{1, 2, 3, 3}

=>{2, 3, 4, 4}

RDD[U] flatMap(f: T => TraversableOnce[U])

TraversableOnce:集合与迭代器的父类

函数将T元素转换为含有新类型U元素的集合,并将这些集合展平(两层转换成一层)后的元素形成新的RDD

rdd.flatMap(x => x.to(3))

{1, 2, 3, 3}

=>{1, 2, 3, 2, 3, 3, 3}

RDD[T] filter(f: T => Boolean)

函数对每个元素进行过滤,通过的元素形成新的RDD

rdd.filter(x => x != 1)

{1, 2, 3, 3}

=>{2, 3, 3}

RDD[T] distinct()

去重

rdd.distinct()

{1, 2, 3, 3}

=>{1, 2, 3}

RDD[U] mapPartitions(f: Iterator[T] => Iterator[U])

map一样,只是转换时是以分区为单位,将一个分区所有元素包装成Iterator一次性传入函数进行处理,而不像map函数那样每个元素都会调用一个函数,即这里有几个分区则才调用几次函数

 

假设有N个元素,有M个分区,那么map的函数的将被调用N,mapPartitions被调用M

val arr = Array(1, 2, 3, 4, 5)

val rdd = sc.parallelize(arr, 2)

rdd.mapPartitions((it: Iterator[Int]) => { var l = List[Int](); it.foreach((e: Int) => l = e * 2 :: l); l.iterator })

=>{2, 4, 6, 8, 10}

RDD[U] mapPartitionsWithIndex(f: (Int, Iterator[T]) => Iterator[U])

mapPartitions类似,不同的时函数多了个分区索引的参数

 

 

RDD[T] union(other: RDD[T])

两个RDD 并集,包括重复的元素

rdd.union(otherRdd)

  { 1, 2, 2, 3, 3}

  { 3, 4, 5}

=>{1, 2, 2, 3, 3, 3, 4, 5}

RDD[T] intersection(other: RDD[T])

两个RDD 交集

rdd.intersection(otherRdd)

  { 1, 2, 2, 3, 3}

  { 3, 4, 5}

=>{3}

RDD[T] subtract(other: RDD[T])

两个RDD相减

rdd.subtract(otherRdd)

  { 1, 2, 2, 3, 3}

  { 3, 4, 5}

=>{1, 2, 2}

RDD[(T, U)] cartesian(other: RDD[U])

两个RDD相减笛卡儿积

rdd.cartesian(otherRdd)

  { 1, 2 }

  { 3, 4}

=>{(1,3),(1,4),(2,3),(2,4)}

RDD[T] sortBy( f: (T) => K, ascending: Boolean,numPartitions: Int)

 

根据转换后的值进行排序,传入的是一个(T) => K 转换函数

rdd.sortBy(_._2, false, 1)

这里根据value进行降序排序

  {("leo", 65), ("tom", 50), ("marry", 100), ("jack", 80)}

=>{("marry", 100),("jack", 80),("leo", 65), ("leo", 65)}

 

RDD[Array[T]] glom()

RDD的每个分区中的类型为T的元素转换换数组Array[T]

 

val arr = Array(1, 2, 3, 4, 5)

val rdd = sc.parallelize(arr, 2)

val arrRDD = rdd.glom()arrRDD.foreach { (arr: Array[Int]) => { println("[ " + arr.mkString(" ") + " ]"); } }

=>[ 1 2 ], [ 3 4 5 ]

 

 

 

 

-RDD转换:

RDD[(K, U)] mapValues[U](f: V => U)

Kkey类型

Vvalue类型

value转换为新的U元素,Key不变

rdd.mapValues(_ + 1)

{"class1", 80), ("class2", 70)}

=>{"class1", 81), ("class2", 71)}

 

RDD[(K, U)] flatMapValues(f: V => TraversableOnce[U])

[K,V]型数据中的Vflatmap操作

rdd.flatMapValues(_.toCharArray())

{ (1, "ab"), (2, "bc")}

=>{(1, 'a'), (1, 'b'), (2, 'b'), (2, 'c')}

RDD[(K, Iterable[V])] groupByKey()

 

根据key进行分组,同一组的元素组成Iterable<V>,并以(key, Iterable<V>)元组类型为元素作为新的RDD返回

rdd.groupByKey()

{("class1", 80), ("class2", 75), ("class1", 90), ("class2", 60)}

=>{("class1",[80,90]),("class2",[75,60])}

 

RDD[(K, Iterable[T])] groupBy(f: T => K)

T:原RDD元素类型

K:新RDD中元素Key的类型

根据函数将元素T映射成相应K后,以此K进行分组

rdd.groupBy({ case 1 => 1; case 2 => 2; case "" => 2 })

  { 1, 2, "" }

=>{(1,[1]),(2,[2, ""])}

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

先根据key进行分组,再对同一组中的的value进行reduce操作:第一次调用函数时传入的是两个Key所对应的value,从第二次往后,传入的两个参数中的第一个为上次函数计算的结果,第二个参数为其它Keyvalue

rdd. reduceByKey(_ + _)

{("class1", 80), ("class2", 75), ("class1", 90), ("class2", 60)}

=>{("class1", 170),("class2", 135)}

RDD[(K, V)] sortByKey()

根据key的大小进行排序(注:并不是先以Key进行分组,再对组类进行排序,而是直接根据Key的值进行排序)

rdd.sortByKey(false)

{(65, "leo"), (50, "tom"),(100, "marry"), (85, "jack")}

=>{(100, "marry"),(85, "jack"),(65, "eo"),(50, "tom")}

comineByKey

 

 

 

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

 

zeroValue:每个分区相同Key累计时的初始值,以及不同分区相同Key合并时的初始值

e.g., Nil for list concatenation, 0 for addition, or 1 for multiplication

对每个value先进行func操作,且funcfoldByKey函数是通过调用comineByKey函数实现的。

zeroVale:对V进行初始化,实际上是通过CombineByKeycreateCombiner实现的  V =>  (zeroValue,V),再通过func函数映射成新的值,即func(zeroValue,V)

 

func: Value将通过func函数按Key值进行合并(实际上是通过CombineByKeymergeValuemergeCombiners函数实现的,只不过在这里,这两个函数是相同的)

val people = List(("Mobin", 1), ("Lucy", 2), ("Amy", 3), ("Amy", 4), ("Lucy", 5))

val rdd = sc.parallelize(people,2)

val foldByKeyRDD = rdd.foldByKey(10)((v1, v2) => { println(v1 + " + " + v2 + " = " + (v1 + v2)); v1 + v2 }) //先对每个V都加10,再对相同Keyvalue值相加

foldByKeyRDD.foreach(println)

//处理第一个分区数据

10 + 1 = 11 // ("Mobin", 1)

10 + 2 = 12 // ("Lucy", 2)

=====================

//处理第二个分区数据

10 + 3 = 13 // ("Amy", 3)

13 + 4 = 17 // ("Amy", 4)同分区同KeyVal先合并

10 + 5 = 15 // ("Lucy", 5)

=====================

//将不同分区相同KeyValue合并起来

12 + 15 = 27 // "Lucy"跨分区,所以需合并

(Amy,17)

(Mobin,11)

(Lucy,27)

RDD[(K, (V, Option[W]))] leftOuterJoin[W](other: RDD[(K, W)]):

左外连接,包含左RDD的所有数据,如果右边没有与之匹配的用None表示

val arr = List(("A", 1), ("A", 2), ("B", 1))

val arr1 = List(("A", "A1"), ("A", "A2"))

val rdd = sc.parallelize(arr, 2)

val rdd1=sc.parallelize(arr1, 2)

val leftOutJoinRDD = rdd.leftOuterJoin(rdd1)

leftOutJoinRDD.foreach(println)

=>

(B,(1,None))

(A,(1,Some(A1)))

(A,(1,Some(A2)))

(A,(2,Some(A1)))

(A,(2,Some(A2)))

RDD[(K, (Option[V], W))] rightOuterJoin[W](other: RDD[(K, W)])

右外连接,包含右RDD的所有数据,如果左边没有与之匹配的用None表示

val arr = List(("A", 1), ("A", 2))

val arr1 = List(("A", "A1"), ("A", "A2"), ("B", 1))

val rdd = sc.parallelize(arr, 2)

val rdd1 = sc.parallelize(arr1, 2)

val leftOutJoinRDD = rdd.rightOuterJoin(rdd1)

leftOutJoinRDD.foreach(println)

(B,(None,1))

(A,(Some(1),A1))

(A,(Some(1),A2))

(A,(Some(2),A1))

(A,(Some(2),A2))

RDD[(K, (V, W))] join(other: RDD[(K, W))

W:另一RDD元素的value的类型

对两个包含<key,value>对的RDD根据key进行join操作,返回类型<key,Tuple2(key,value)>

rdd.join(otherRdd)

  {(1, "leo"),(2, "jack"),(3, "tom")}

  {(1, 100), (2, 90), (3, 60), (1, 70), (2, 80), (3, 50)}

=>{(1,("leo",100)),(1,("leo",70)),(2, ("jack",90),(2, ("jack",80),(3, ("tom",60),(3, ("tom",50))}

RDD[(K, (Iterable[V], Iterable[W]))] cogroup(other: RDD[(K, W)])

join,也是根据key进行join,只不过相同keyvalue分别存放到Iterable<value>

rdd.cogroup(otherRdd)

  {(1, "leo"),(2, "jack"),(3, "tom")}

  {(1, 100), (2, 90), (3, 60), (1, 70), (2, 80), (3, 50)}

=>{(1,(["leo"],[100,70])),(2, (["jack"],[90,80])),(3, (["tom","lily"],[60,50]))}

 

 

 

 

常用Action

T reduce(f: (T, T) => T)

对所有元素进行reduce操作

rdd.reduce(_ + _)

{1, 2, 2, 3, 3, 3}

=>14

Array[T] collect()

RDD中所有元素返回到一个数组里

注意:This method should only be used if the resulting array is expected to be small, as all the data is loaded into the driver's memory.

rdd.collect()

{1, 2, 3, 3}

=>[1, 2, 3, 3]

Map[K, V] collectAsMap()

作用于K-V类型的RDD上,作用与collect不同的是collectAsMap函数不包含重复的key,对于重复的key,后面的元素覆盖前面的元素

rdd.collectAsMap()

{ ("leo", 65), ("tom", 50), ("tom", 100)}

=>{ ("leo", 65), ("tom", 100)}

Long count()

统计RDD 中的元素个数

rdd.count()

{1, 2, 3, 3}

=>4

Map[T, Long] countByValue()

各元素在 RDD 中出现的次数

注意:This method should only be used if the resulting map is expected to be small, as the whole thing is loaded into the driver's memory.

To handle very large results, consider using rdd.map(x => (x, 1L)).reduceByKey(_ + _), which returns an RDD[T, Long] instead of a map.

rdd.countByValue()

{1, 2, 3, 3}

=>Map(1 -> 1, 3 -> 2, 2 -> 1)

Map[K, Long] countByKey()

根据Key进行分组,再对每组里的value分别进行计数统计

注意:This method should only be used if the resulting map is expected to be small, as the whole thing is loaded into the driver's memory.

 

To handle very large results, consider using rdd.mapValues(_ => 1L).reduceByKey(_ + _), which returns an RDD[T, Long] instead of a map.

 

  { ("leo", 65), ("tom", 50), ("tom", 100), ("tom", 100) }

=>Map(leo -> 1, tom -> 3)

T first()

取第一个元素,实质上是调用take(1)实现的

rdd.first()

{3, 2, 1, 4}

=>3

Array[T] take(num: Int)

RDD 中返回前 num 个元素

注意:This method should only be used if the resulting array is expected to be small, as all the data is loaded into the driver's memory.

rdd.take(2)

 

  {3, 2, 1, 4}

=>[3, 2]

Array[T] top(num: Int ) (implicit ord: Ordering[T])

 

如果没有传递 ord参数,则使用隐式参数,且提供的默认隐式参数为升序排序,可以传递一个自定义的Ordering来覆盖默认提供。 top实现是将Ordering反序后再调用  takeOrdered的:takeOrdered(num)(ord.reverse)

默认从 RDD 中返回最最大的 num个元素

注意:This method should only be used if the resulting array is expected to be small, as all the data is loaded into the driver's memory.

rdd.top(2)

  {3, 2, 1, 4}

=>[4, 3]

Array[T] takeOrdered(num: Int)(implicit ord: Ordering[T])

 

如果没有传递 ord参数,则使用隐式参数,且提供的默认隐式参数为升序排序,可以传递一个自定义的Ordering来覆盖默认提供

top相反,默认取的是前面最小的num个元素

注意:This method should only be used if the resulting array is expected to be small, as all the data is loaded into the driver's memory.

rdd.takeOrdered(2)(myOrdering)

  {3, 2, 1, 4}

=>[1, 2]

T fold(zeroValue: T)(op: (T, T) => T)

 

zeroValue:为每个分区累计的初始值,以及不同分区累计的初始值

e.g., Nil for list concatenation, 0 for addition, or 1 for multiplication

reduce() 样,

提供初始值。注意:每个分区应用op函数时,都会以zeroValue为初始值进行计算,然后将每个分区的结果合并时,还是会以zeroValue为初始值进行合并计算

val arr = Array(1, 2, 3, 4, 5);

val rdd = sc.parallelize(arr, 2) //分成两分区[1, 2] [3, 4, 5]

println(rdd.fold(10)((v1, v2) => { println(v1 + " + " + v2 + " = " + (v1 + v2)); v1 + v2 }))

//处理第一个分区数据

10 + 1 = 11

11 + 2 = 13 //从第二个元素起,每分区内先累加

=====================

//处理第一个分区数据

10 + 3 = 13

13 + 4 = 17 //从第二个元素起,每分区内先累加

17 + 5 = 22 //从第二个元素起,每分区内先累加

=====================

//将各分区汇总起来

10 + 13 = 23 // 汇总时还会使用初始值来作起始

23 + 22 = 45

45

 

 

 

 

U aggregate (zeroValue: U)(seqOp: (U, T) => U, combOp: (U, U) => U)

 

初始值类型与原始数据类型可以不同,但初始值类型决定了返回值类型

fold一样,计算时需要提供初始值,不同的是,分区的计算函数(seqOp)与分区合并计算函数(combOp)是不同的,但fold分区计算函数与分区合并计算函数是同一函数

rdd.fold(5)(_ + _, _ + _)

val arr = Array(1, 2, 3, 4);

val rdd = sc.parallelize(arr, 2)

println(rdd.aggregate(5)(

(v1, v2) => { println("v1 = " + v1 + " ; v2 = " + v2); v1 + v2 },

(v1, v2) => { println("v1 = " + v1 + " ; v2 = " + v2); v1 + v2 })

)

过程与结果与上面的fold函数一样

Unit saveAsTextFile(path: String)

RDD元素保存到文件中,对每个元素调用toString方法

 

 

Unit foreach(f: T => Unit)

遍历RDD中的每个元素

rdd.foreach(println(_))

 

 

 

 

 

 

 

 

 

comineByKey

  def combineByKey[C](

      createCombiner: V => C,

      mergeValue: (C, V) => C,

      mergeCombiners: (C, C) => C,

      partitioner: Partitioner,

      mapSideCombine: Boolean = true,

      serializer: Serializer = null): RDD[(K, C)]

createCombiner:在第一次遇到Key时创建组合器函数,将RDD数据集中的V类型值转换C类型值(V => C),

mergeValue:合并值函数,再次遇到相同的Key时,将createCombiner道理的C类型值与这次传入的V类型值合并成一个C类型值(C,V=>C

mergeCombiners:合并组合器函数,将C类型值两两合并成一个C类型值

partitioner:使用已有的或自定义的分区函数,默认是HashPartitioner

mapSideCombine:是否在map端进行Combine操作,默认为true

 

例:统计男性和女生的个数,并以(性别,(名字,名字....),个数)的形式输出

object CombineByKey {

  def main(args: Array[String]) {

    val conf = new SparkConf().setMaster("local").setAppName("combinByKey")

    val sc = new SparkContext(conf)

    val people = List(("male", "Mobin"), ("male", "Kpop"), ("female", "Lucy"), ("male", "Lufei"), ("female", "Amy"))

    val rdd = sc.parallelize(people)

    val combinByKeyRDD = rdd.combineByKey(

      (x: String) => (List(x), 1),

      (peo: (List[String], Int), x: String) => (x :: peo._1, peo._2 + 1),

      (sex1: (List[String], Int), sex2: (List[String], Int)) => (sex1._1 ::: sex2._1, sex1._2 + sex2._2))

    combinByKeyRDD.foreach(println)

    sc.stop()

  }

}

输出:

(male,(List(Lufei, Kpop, Mobin),3))

(female,(List(Amy, Lucy),2))

 

计算过程:

Partition1:

K="male"  -->  ("male","Mobin")  --> createCombiner("Mobin") =>  peo1 = (  List("Mobin") , 1 )

K="male"  -->  ("male","Kpop")  --> mergeValue(peo1,"Kpop") =>  peo2 = (  "Kpop"  ::  peo1_1 , 1 + 1 )    //Key相同调用mergeValue函数对值进行合并

K="female"  -->  ("female","Lucy")  --> createCombiner("Lucy") =>  peo3 = (  List("Lucy") , 1 )

 

Partition2:

K="male"  -->  ("male","Lufei")  --> createCombiner("Lufei") =>  peo4 = (  List("Lufei") , 1 )

K="female"  -->  ("female","Amy")  --> createCombiner("Amy") =>  peo5 = (  List("Amy") , 1 )

 

Merger Partition:

K="male" --> mergeCombiners(peo2,peo4) => (List(Lufei,Kpop,Mobin))

K="female" --> mergeCombiners(peo3,peo5) => (List(Amy,Lucy))

 

 

 

 

附件列表

     

    posted @ 2017-07-31 12:42  江正军  阅读(1697)  评论(0编辑  收藏  举报