4.1 Transformation算子
基本初始化
private val conf: SparkConf = new SparkConf().setAppName("TestTransformation").setMaster("local")
private val sparkContext = new SparkContext(conf)
4.2 map、flatMap、mapParations、mapPartitionsWithIndex
Map
代码:
def map(): Unit ={
val list = List("张无忌", "赵敏", "周芷若")
val listRDD = sc.parallelize(list)
val nameRDD = listRDD.map(name => "Hello " + name)
nameRDD.foreach(name => println(name))
}
运行结果:
总结:
可以看出,对于map算子,源listRDD的每个元素都会进行计算,由于是依次进行传参,所以他是有序的,新RDD的元素顺序与源RDD是相同的。而由有序又引出接下来的flatMap。
Flatmap
代码:
def flatMap(): Unit ={
val list = List("张无忌 赵敏","宋青书 周芷若")
val listRDD = sc.parallelize(list)
val nameRDD = listRDD.flatMap(line => line.split(" ")).map(name => "Hello " + name)
nameRDD.foreach(name => println(name))
}
运行结果:
总结:
flatMap的特性决定了这个算子在对需要随时增加元素的时候十分好用,比如在对源RDD查漏补缺时。
map和flatMap都是依次进行参数传递的,但有时候需要RDD中的两个元素进行相应操作时(例如:算存款所得时,下一个月所得的利息是要原本金加上上一个月所得的本金的),这两个算子便无法达到目的了,这是便需要mapPartitions算子,他传参的方式是将整个RDD传入,然后将一个迭代器传出生成一个新的RDD,由于整个RDD都传入了,所以便能完成前面说的业务。
mapPartitions
代码:
def mapParations(): Unit ={
val list = List(1,2,3,4,5,6)
val listRDD = sc.parallelize(list,2)
listRDD.mapPartitions(iterator => {
val newList: ListBuffer[String] = ListBuffer()
while (iterator.hasNext){
newList.append("hello " + iterator.next())
}
newList.toIterator
}).foreach(name => println(name))
}
运行结果:
mapPartitionsWithIndex
每次获取和处理的就是一个分区的数据,并且知道处理的分区的分区号是啥?
代码:
def mapPartitionsWithIndex(): Unit ={
val list = List(1,2,3,4,5,6,7,8)
sc.parallelize(list).mapPartitionsWithIndex((index,iterator) => {
val listBuffer:ListBuffer[String] = new ListBuffer
while (iterator.hasNext){
listBuffer.append(index+"_"+iterator.next())
}
listBuffer.iterator
},true)
.foreach(println(_))
}
运行结果:
4.3 reduce、reduceByKey
Reduce
reduce其实是将RDD中的所有元素进行合并,当运行call方法时,会传入两个参数,在call方法中将两个参数合并后返回,而这个返回值回合一个新的RDD中的元素再次传入call方法中,继续合并,直到合并到只剩下一个元素时。
代码:
def reduce(): Unit ={
val list = List(1,2,3,4,5,6)
val listRDD = sc.parallelize(list)
val result = listRDD.reduce((x,y) => x+y)
println(result)
}
运行结果:
reduceByKey
reduceByKey仅将RDD中所有K,V对中K值相同的V进行合并。
代码:
def reduceByKey(): Unit ={
val list = List(("武当", 99), ("少林", 97), ("武当", 89), ("少林", 77))
val mapRDD = sc.parallelize(list)
val resultRDD = mapRDD.reduceByKey(_+_)
resultRDD.foreach(tuple => println("门派: " + tuple._1 + "->" + tuple._2))
}
运行结果:
4.4 union,join和groupByKey
Union
当要将两个RDD合并时,便要用到union和join,其中union只是简单的将两个RDD累加起来,可以看做List的addAll方法。就想List中一样,当使用union及join时,必须保证两个RDD的泛型是一致的。
代码:
def union(): Unit ={
val list1 = List(1,2,3,4)
val list2 = List(3,4,5,6)
val rdd1 = sc.parallelize(list1)
val rdd2 = sc.parallelize(list2)
rdd1.union(rdd2).foreach(println(_))
}
运行结果:
groupByKey
在说join之前说说groupByKey,因为join可以理解为union与groupByKey的结合:groupBy是将RDD中的元素进行分组,组名是call方法中的返回值,而顾名思义groupByKey是将PairRDD中拥有相同key值得元素归为一组。即:
代码:
def groupByKey(): Unit ={
val list = List(("武当", "张三丰"), ("峨眉", "灭绝师太"), ("武当", "宋青书"), ("峨眉", "周芷若"))
val listRDD = sc.parallelize(list)
val groupByKeyRDD = listRDD.groupByKey()
groupByKeyRDD.foreach(t => {
val menpai = t._1
val iterator = t._2.iterator
var people = ""
while (iterator.hasNext) people = people + iterator.next + " "
println("门派:" + menpai + "人员:" + people)
})
}
运行结果:
Join
join是将两个PairRDD合并,并将有相同key的元素分为一组,可以理解为groupByKey和Union的结合
代码:
def join(): Unit = {
val list1 = List((1, "东方不败"), (2, "令狐冲"), (3, "林平之"))
val list2 = List((1, 99), (2, 98), (3, 97))
val list1RDD = sc.parallelize(list1)
val list2RDD = sc.parallelize(list2)
val joinRDD = list1RDD.join(list2RDD)
joinRDD.foreach(t => println("学号:" + t._1 + " 姓名:" + t._2._1 + " 成绩:" + t._2._2))
}
运行结果:
4.5 sample、cartesian
Sample
Sample(withReplacement : scala.Boolean, fraction : scala.Double,seed scala.Long)
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,01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
代码:
def sample(): Unit ={
val list = 1 to 100
val listRDD = sc.parallelize(list)
listRDD.sample(false,0.1,0).foreach(num => print(num + " "))
}
运行结果:
Cartesian
cartesian是用于求笛卡尔积的
代码:
def cartesian(): Unit ={
val list1 = List("A","B")
val list2 = List(1,2,3)
val list1RDD = sc.parallelize(list1)
val list2RDD = sc.parallelize(list2)
list1RDD.cartesian(list2RDD).foreach(t => println(t._1 +"->"+t._2))
}
运行结果:
4.6 filter、distinct、intersection
Filter
过滤出偶数
代码:
def filter(): Unit ={
val list = List(1,2,3,4,5,6,7,8,9,10)
val listRDD = sc.parallelize(list)
listRDD.filter(num => num % 2 ==0).foreach(print(_))
}
运行结果:
Distinct
去重
代码:
def distinct(): Unit ={
val list = List(1,1,2,2,3,3,4,5)
sc.parallelize(list).distinct().foreach(println(_))
}
运行结果:
Intersection
代码:
def intersection(): Unit ={
val list1 = List(1,2,3,4)
val list2 = List(3,4,5,6)
val list1RDD = sc.parallelize(list1)
val list2RDD = sc.parallelize(list2)
list1RDD.intersection(list2RDD).foreach(println(_))
}
运行结果:
4.7 coalesce、repartition、repartitionAndSortWithinPartitions
Coalesce
分区数由多 -》 变少
代码:
def coalesce(): Unit = {
val list = List(1,2,3,4,5,6,7,8,9)
sc.parallelize(list,3).coalesce(1).foreach(println(_))
}
运行结果:
Replication
进行重分区,解决的问题:本来分区数少 -》 增加分区数
代码:
def replication(): Unit ={
val list = List(1,2,3,4)
val listRDD = sc.parallelize(list,1)
listRDD.repartition(2).foreach(println(_))
}
运行结果:
repartitionAndSortWithinPartitions
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。
代码:
def repartitionAndSortWithinPartitions(): Unit ={
val list = List(1, 4, 55, 66, 33, 48, 23)
val listRDD = sc.parallelize(list,1)
listRDD.map(num => (num,num))
.repartitionAndSortWithinPartitions(new HashPartitioner(2))
.mapPartitionsWithIndex((index,iterator) => {
val listBuffer: ListBuffer[String] = new ListBuffer
while (iterator.hasNext) {
listBuffer.append(index + "_" + iterator.next())
}
listBuffer.iterator
},false)
.foreach(println(_))
}
运行结果:
4.8 cogroup、sortBykey、aggregateByKey
Cogroup
对两个RDD中的KV元素,每个RDD中相同key中的元素分别聚合成一个集合。与reduceByKey不同的是针对两个RDD中相同的key的元素进行合并。
代码:
def cogroup(): Unit ={
val list1 = List((1, "www"), (2, "bbs"))
val list2 = List((1, "cnblog"), (2, "cnblog"), (3, "very"))
val list3 = List((1, "com"), (2, "com"), (3, "good"))
val list1RDD = sc.parallelize(list1)
val list2RDD = sc.parallelize(list2)
val list3RDD = sc.parallelize(list3)
list1RDD.cogroup(list2RDD,list3RDD).foreach(tuple =>
println(tuple._1 + " " + tuple._2._1 + " " + tuple._2._2 + " " + tuple._2._3))
}
运行结果:
sortBykey
sortByKey函数作用于Key-Value形式的RDD,并对Key进行排序。它是在org.apache.spark.rdd.OrderedRDDFunctions中实现的,实现如下
def sortByKey(ascending: Boolean = true, numPartitions: Int = self.partitions.size)
: RDD[(K, V)] =
{
val part = new RangePartitioner(numPartitions, self, ascending)
new ShuffledRDD[K, V, V](self, part)
.setKeyOrdering(if (ascending) ordering else ordering.reverse)
}
从函数的实现可以看出,它主要接受两个函数,含义和sortBy一样,这里就不进行解释了。该函数返回的RDD一定是ShuffledRDD类型的,因为对源RDD进行排序,必须进行Shuffle操作,而Shuffle操作的结果RDD就是ShuffledRDD。其实这个函数的实现很优雅,里面用到了RangePartitioner,它可以使得相应的范围Key数据分到同一个partition中,然后内部用到了mapPartitions对每个partition中的数据进行排序,而每个partition中数据的排序用到了标准的sort机制,避免了大量数据的shuffle。下面对sortByKey的使用进行说明:
代码:
def sortByKey(): Unit ={
val list = List((99, "张三丰"), (96, "东方不败"), (66, "林平之"), (98, "聂风"))
sc.parallelize(list).sortByKey(false).foreach(tuple => println(tuple._2 + "->" + tuple._1))
}
运行结果:
aggregateByKey
aggregateByKey函数对PairRDD中相同Key的值进行聚合操作,在聚合过程中同样使用了一个中立的初始值。和aggregate函数类似,aggregateByKey返回值的类型不需要和RDD中value的类型一致。因为aggregateByKey是对相同Key中的值进行聚合操作,所以aggregateByKey函数最终返回的类型还是Pair RDD,对应的结果是Key和聚合好的值;而aggregate函数直接是返回非RDD的结果,这点需要注意。在实现过程中,定义了三个aggregateByKey函数原型,但最终调用的aggregateByKey函数都一致。
代码:
def aggregateByKey(): Unit ={
val list = List("you,jump", "i,jump")
sc.parallelize(list)
.flatMap(_.split(","))
.map((_, 1))
.aggregateByKey(0)(_+_,_+_)
.foreach(tuple =>println(tuple._1+"->"+tuple._2))
}
运行结果: