1.map 和 mapPartitions
map的输入变换函数应用于RDD中所有元素,而mapPartitions应用于所有分区。区别于mapPartitions主要在于调用粒度不同。
mapPartition可以倒过来理解,先partition,再把每个partition进行map函数,
适用场景:
如果在映射的过程中需要频繁创建额外的对象,使用mapPartitions要比map高效的多。
创新互联公司专注于正宁网站建设服务及定制,我们拥有丰富的企业做网站经验。 热诚为您提供正宁营销型网站建设,正宁网站制作、正宁网页设计、正宁网站官网定制、微信平台小程序开发服务,打造正宁网络公司原创品牌,更为您提供正宁网站排名全网营销落地服务。
val numbers: RDD[Int] = sc.parallelize(seqs,3)
//map
numbers.map(x => {
println("AAA")//打印6次
x * 3
}).collect().foreach(println(_))
/**
* 遍历分区(3个)
*/
numbers.mapPartitions(par => {
println("aaa")//打印3次
par.map(p => p * 3)
}).collect().foreach(println(_))
2.filter
过滤操作,满足filter内function函数为true的RDD内所有元素组成一个新的数据集
val seqs = Seq(1,2,3,4,5,6)
//4,5,6
seqs.filter(x=> x > 3).foreach(println(_))
3.flatMap
map是对RDD元素逐一进行函数操作映射为另外一个RDD,
而flatMap操作是将函数应用于RDD之中的每一个元素,将
返回迭代器的所有内容构成的新的RDD。
flatMap和Map区别在于map为“映射”,而flatMap则是“先映射,后扁平化”。
val seqs = Array("aaa AAA","bbb BBB","ccc CCC","ddd DDD")
val numbers = sc.parallelize(seqs)
scala> numbers.map(x => x.split(" ")).collect()
res1: Array[Array[String]] = Array(Array(aaa, AAA), Array(bbb, BBB), Array(ccc, CCC), Array(ddd, DDD))
scala> numbers.flatMap(x=>x.split(" ")).collect()
res2: Array[String] = Array(aaa, AAA, bbb, BBB, ccc, CCC, ddd, DDD)
4.mapPartitionsWithIndex
与mapPartitions类似,但需要提供一个表示分区索引值的整型值作为参数,因此function必须是(int, Iterator\
//统计键值对中的各个分区的元素
val rdd = sc.parallelize(List((1,1), (1,2), (2,3), (2,4), (3,5), (3,6),(4,7), (4,8),(5,9), (5,10)),3)
def mapPartIndexFunc(i1:Int,iter: Iterator[(Int,Int)]):Iterator[(Int,(Int,Int))]={
var res = List[(Int,(Int,Int))]()
while(iter.hasNext){
var next = iter.next()
res=res.::(i1,next)
}
res.iterator
}
val mapPartIndexRDD = rdd.mapPartitionsWithIndex(mapPartIndexFunc)
mapPartIndexRDD.foreach(println(_))
//计算结果
(0,(1,1))
(0,(1,2))
(0,(2,3))
(1,(2,4))
(1,(3,5))
(1,(3,6))
(2,(4,7))
(2,(4,8))
(2,(5,9))
(2,(5,10))
5.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
scala> val rdd = sc.parallelize(1 to 10)
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[9] at parallelize at :24
scala> rdd.sample(true,0.57,5).collect
res10: Array[Int] = Array(8, 8, 8, 9)
6.union(并集)
合并,它只是将rdd1和rdd2在逻辑上合并,并不会进行数据的合并以传输,不去重
scala>var rdd1 = sc.parallelize(List("aa","aa","bb","cc","dd"));
rdd1: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[4] at parallelize at :24
scala>var rdd2 = sc.parallelize(List("aa","dd","ff"));
rdd2: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[5] at parallelize at :24
scala>rdd1.union(rdd2).collect();
res3: Array[String] = Array(aa, aa, bb, cc, dd, aa, dd, ff)
7.intersection
RDD1.intersection(RDD2),返回两个RDD的交集,并且去重
intersection需要混洗数据,比较浪费性能
scala> var RDD1 = sc.parallelize(List("aa","aa","bb","cc","dd"))
RDD1: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[11] at parallelize at :24
scala> var RDD2 = sc.parallelize(List("aa","dd","ff"))
RDD2: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[12] at parallelize at :24
scala> RDD1.intersection(RDD2).collect
res5: Array[String] = Array(aa, dd)
8.distinct
distinct用于去重, 我们生成的RDD可能有重复的元素,使用distinct方法可以去掉重复的元素, 不过此方法涉及到混洗,操作开销很大
scala> var RDD1 = sc.parallelize(List("aa","aa","bb","cc","dd"))
RDD1: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[7] at parallelize at :24
scala> RDD1.collect
res4: Array[String] = Array(aa, aa, bb, cc, dd)
scala> val distinctRDD = RDD1.distinct.collect
distinctRDD: Array[String] = Array(aa, dd, bb, cc)
9.groupByKey
groupByKey会将RDD[key,value] 按照相同的key进行分组,形成RDD[key,Iterable[value]]的形式, 有点类似于sql中的groupby,例如类似于MySQL中的group_concat
//按照学生姓名对学生成绩进行分组
scala> val scoreDetail = sc.parallelize(List(("xiaoming",75),("xiaoming",90),("lihua",95),("lihua",100),("xiaofeng",85)))
scoreDetail: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[3] at parallelize at :24
scala> scoreDetail.groupByKey().collect().foreach(println(_))
(lihua,CompactBuffer(95, 100))
(xiaoming,CompactBuffer(75, 90))
(xiaofeng,CompactBuffer(85))
10.reduceByKey
接收一个函数,按照相同的key进行reduce操作,类似于scala的reduce的操作
例如RDD {(1, 2), (3, 4), (3, 6)}进行reduce ,key不变,value相加
scala> var mapRDD = sc.parallelize(List((1,2),(3,4),(3,6)))
mapRDD: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[6] at parallelize at :24
scala> var reduceRDD = mapRDD.reduceByKey(_+_)
reduceRDD: org.apache.spark.rdd.RDD[(Int, Int)] = ShuffledRDD[7] at reduceByKey at :26
scala> reduceRDD.foreach(x=>println(x))
(1,2)
(3,10)
11.aggregateByKey
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,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
val data = sc.parallelize(List((1,3),(1,2),(1,4),(2,3),(3,7),(3,8)),1)
//println(data.partitions.size)
/**
* 比较相同key得两个value中的最大值,第一次为max(1,3),1为初始值,得:3,第二次为max(3,2),得3,
* 第三次为max(3,4),得:4,所以key为1的结果为:(1,4)
*/
def seq(a:Int, b:Int) : Int ={
math.max(a,b)
}
/**
* 不同分区中相同key的value相加,如果只有一个分区,此方法不起效果
*/
def comb(a:Int, b:Int) : Int ={
a + b
}
//聚合打印结果
data.aggregateByKey(1)(seq, comb).collect.foreach(println(_))
//查看各个分区数据
data.mapPartitionsWithIndex {
(partid, iter) => {
var part_map = scala.collection.mutable.Map[String, List[(Int,Int)]]()
var part_name = "part_" + partid
part_map(part_name) = List[(Int,Int)]()
while (iter.hasNext) {
part_map(part_name) :+= iter.next() //:+= 列表尾部追加元素
}
part_map.iterator
}
}.collect().foreach(println(_))
12.sortByKey
用于对pairRDD按照key进行排序,第一个参数可以设置true或者false,默认是true
scala> val rdd = sc.parallelize(Array((3, 4),(1, 2),(4,4),(2,5), (6,5), (5, 6)))
rdd: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[10] at parallelize at :24
scala> rdd.sortByKey().collect
res4: Array[(Int, Int)] = Array((1,2), (2,5), (3,4), (4,4), (5,6), (6,5))
scala> rdd.sortByKey(true).collect
res5: Array[(Int, Int)] = Array((1,2), (2,5), (3,4), (4,4), (5,6), (6,5))
scala> rdd.sortByKey(false).collect
res6: Array[(Int, Int)] = Array((6,5), (5,6), (4,4), (3,4), (2,5), (1,2))
13.join
RDD1.join(RDD2) ,可以把RDD1,RDD2中的相同的key给连接起来,类似于sql中的inner join操作,返回两边都匹配的数据
scala> val RDD1 = sc.parallelize(Array(("A","a1"),("B","b1"),("C","c1"),("D","d1"),("E","e1"),("F","f1")))
RDD1: org.apache.spark.rdd.RDD[(String, String)] = ParallelCollectionRDD[19] at parallelize at :24
scala> val RDD2 = sc.parallelize(Array(("A","a2"),("B","b2"),("C","c1"),("C","c2"),("C","c3"),("E","e2")))
RDD2: org.apache.spark.rdd.RDD[(String, String)] = ParallelCollectionRDD[20] at parallelize at :24
scala> RDD1.join(RDD2).collect
res8: Array[(String, (String, String))] = Array((B,(b1,b2)), (A,(a1,a2)), (C,(c1,c1)), (C,(c1,c2)), (C,(c1,c3)), (E,(e1,e2)))
scala> RDD2.join(RDD1).collect
res9: Array[(String, (String, String))] = Array((B,(b2,b1)), (A,(a2,a1)), (C,(c1,c1)), (C,(c2,c1)), (C,(c3,c1)), (E,(e2,e1)))
其他操作:
left outer join:是以左边为基准,向左靠(左边(a)的记录一定会存在,右边(b)的记录有的返回Some(x),没有的补None。)
scala> RDD1.leftOuterJoin(RDD2).collect
res11: Array[(String, (String, Option[String]))] = Array((B,(b1,Some(b2))), (A,(a1,Some(a2))), (C,(c1,Some(c1))), (C,(c1,Some(c2))), (C,(c1,Some(c3))), (E,(e1,Some(e2))), (F,(f1,None)), (D,(d1,None)))
scala> RDD2.leftOuterJoin(RDD1).collect
res12: Array[(String, (String, Option[String]))] = Array((B,(b2,Some(b1))), (A,(a2,Some(a1))), (C,(c1,Some(c1))), (C,(c2,Some(c1))), (C,(c3,Some(c1))), (E,(e2,Some(e1))))
right outer join:是以右边为基准,向右靠(右边(b)的记录一定会存在,左边(a)的记录有的返回Some(x),没有的补None。)
scala> RDD1.rightOuterJoin(RDD2).collect
res13: Array[(String, (Option[String], String))] = Array((B,(Some(b1),b2)), (A,(Some(a1),a2)), (C,(Some(c1),c1)), (C,(Some(c1),c2)), (C,(Some(c1),c3)), (E,(Some(e1),e2)))
scala> RDD2.rightOuterJoin(RDD1).collect
res14: Array[(String, (Option[String], String))] = Array((B,(Some(b2),b1)), (A,(Some(a2),a1)), (C,(Some(c1),c1)), (C,(Some(c2),c1)), (C,(Some(c3),c1)), (E,(Some(e2),e1)), (F,(None,f1)), (D,(None,d1)))
full outer join:左边和右边的都一定存在
scala> RDD1.fullOuterJoin(RDD2).collect
res16: Array[(String, (Option[String], Option[String]))] = Array((B,(Some(b1),Some(b2))), (A,(Some(a1),Some(a2))), (C,(Some(c1),Some(c1))), (C,(Some(c1),Some(c2))), (C,(Some(c1),Some(c3))), (E,(Some(e1),Some(e2))), (F,(Some(f1),None)), (D,(Some(d1),None)))
scala> RDD2.fullOuterJoin(RDD1).collect
res17: Array[(String, (Option[String], Option[String]))] = Array((B,(Some(b2),Some(b1))), (A,(Some(a2),Some(a1))), (C,(Some(c1),Some(c1))), (C,(Some(c2),Some(c1))), (C,(Some(c3),Some(c1))), (E,(Some(e2),Some(e1))), (F,(None,Some(f1))), (D,(None,Some(d1))))
14.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。
例子中将多个RDD中同一个Key对应的Value组合到一起。rdd1中不存在Key为dd的元素(自然就不存在Value了),在组合的过程中将rdd1对应的位置
设置为CompactBuffer()了,而不是去掉了。
scala> val rdd1 = sc.parallelize(Array(("aa",1),("bb",2),("cc",6)))
rdd1: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[6] at parallelize at :24
scala> val rdd2 = sc.parallelize(Array(("aa",3),("bb",4),("cc",5),("dd",6),("aa",8)))
rdd2: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[7] at parallelize at :24
scala> val rdd3 = rdd1.cogroup(rdd2).collect
rdd3: Array[(String, (Iterable[Int], Iterable[Int]))] = Array((aa,(CompactBuffer(1),CompactBuffer(3, 8))), (dd,(CompactBuffer(),CompactBuffer(6))), (bb,(CompactBuffer(2),CompactBuffer(4))), (cc,(CompactBuffer(6),CompactBuffer(5))))
15.cartesian(笛卡尔积)
RDD1.cartesian(RDD2) 返回RDD1和RDD2的笛卡儿积,这个开销非常大
scala> var RDD1 = sc.parallelize(List("1","2","3"))
RDD1: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[0] at parallelize at :24
scala> var RDD2 = sc.parallelize(List("a","b","c"))
RDD2: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[1] at parallelize at :24
scala> RDD1.cartesian(RDD2).collect
res0: Array[(String, String)] = Array((1,a), (1,b), (1,c), (2,a), (2,b), (2,c), (3,a), (3,b), (3,c))
16.pipe
有种特殊的Rdd,即pipedRdd,提供了调用外部程序如基于CUDA的C++程序,使其能够更快的进行计算。caffe on spark 和tensorflow on spark 也是基于此机制
#准备脚本
#!/bin/sh
echo "Running shell script"
while read LINE; do
echo ${LINE}!
done
# Spark RDD调用
scala> val data = sc.parallelize(List("hi","hello","how","are","you"))
data: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[52] at parallelize at :24
scala> val scriptPath = "/home/hadoop/echo.sh"
scriptPath: String = /home/hadoop/echo.sh
scala> val pipeRDD = data.pipe(scriptPath)
pipeRDD: org.apache.spark.rdd.RDD[String] = PipedRDD[53] at pipe at :28
scala> pipeRDD.collect()
res21: Array[String] = Array(Running shell script, hi!, hello!, how!, are!, you!)
17.coalesce 和 repartition
他们两个都是RDD的分区进行重新划分,repartition只是coalesce接口中shuffle为true的简易实现。
repartition一定会发生shuffle过程
coalesce则不一定
def repartition(numPartitions: Int)(implicit ord: Ordering[T] = null): RDD[T] = withScope {
coalesce(numPartitions, shuffle = true)
}
假设RDD有N个分区,需要重新划分成M个分区
1)NM并且N和M相差不多,(假如N是1000,M是100)那么就可以将N个分区中的若干个分区合并成一个新的分区,最终合并为M个分区,这时可以将shuff设置为false,在shuffl为false的情况下,如果M>N时,coalesce为无效的,不进行shuffle过程,父RDD和子RDD之间是窄依赖关系。
3)如果N>M并且两者相差悬殊,这时如果将shuffle设置为false,父子RDD是窄依赖关系,他们同处在一个Stage中,就可能造成spark程序的并行度不够,从而影响性能,如果在M为1的时候,为了使coalesce之前的操作有更好的并行度,可以讲shuffle设置为true。
总之:如果shuff为false时,如果传入的参数大于现有的分区数目,RDD的分区数不变,也就是说不经过shuffle,是无法将RDDde分区数变多的。
18.repartitionAndSortWithinPartitions
根据给定的分区程序对RDD进行重新分区,并在每个生成的分区内按键对记录进行排序。 这比调用重新分区,然后在每个分区内进行排序更有效率,因为它可以将排序压入洗牌机器。
repartitionAndSortWithinPartitions算是一个高效的算子,是因为它要比使用repartition And sortByKey 效率高,这是由于它的排序是在shuffle过程中进行,一边shuffle,一边排序;
package core
import org.apache.spark.rdd.RDD
import org.apache.spark.{HashPartitioner, SparkConf, SparkContext}
object TransformationDemo {
def main(args: Array[String]): Unit = {
val sparkConf: SparkConf = new SparkConf().setMaster("local").setAppName("test")
val sc = new SparkContext(sparkConf)
val rdd1: RDD[(Int, Int)] = sc.parallelize(List((1,2),(2,3),(3,7),(4,8),(5,2),(6,5),(7,7)))
//1
println(rdd1.partitions.size)
/**
* (0,(7,7))
* (0,(6,5))
* (0,(5,2))
* (0,(4,8))
* (0,(3,7))
* (0,(2,3))
* (0,(1,2))
*/
rdd1.mapPartitionsWithIndex(mapPartIndexFunc).foreach(println)
//重新分区并排序(默认根据key升序排序)
val rdd2: RDD[(Int, Int)] = rdd1.repartitionAndSortWithinPartitions(new HashPartitioner(3))
//3
println(rdd2.partitions.size)
/**
* (0,(6,5))
* (0,(3,7))
* (1,(7,7))
* (1,(4,8))
* (1,(1,2))
* (2,(5,2))
* (2,(2,3))
*/
rdd2.mapPartitionsWithIndex(mapPartIndexFunc).foreach(println)
/**
* (3,7)
* (6,5)
* (1,2)
* (4,8)
* (7,7)
* (2,3)
* (5,2)
*/
rdd2.collect().foreach(println)
sc.stop()
}
/**
* 遍历获取每个分区中的数据
* @param i1
* @param iter
* @return
*/
def mapPartIndexFunc(i1:Int,iter: Iterator[(Int,Int)]):Iterator[(Int,(Int,Int))]={
var res = List[(Int,(Int,Int))]()
while(iter.hasNext){
var next = iter.next()
res=res.::(i1,next)
}
res.iterator
}
}
网站标题:Spark常用的transformation算子
转载来源:http://scyingshan.cn/article/ggcige.html