摘要:
RDD:弹性分布式数据集,是一种特殊集合 ‚ 支持多种来源 ‚ 有容错机制 ‚ 可以被缓存 ‚ 支持并行操作,一个RDD代表一个分区里的数据集
RDD有两种操作算子:
Transformation(转换):Transformation属于延迟计算,当一个RDD转换成另一个RDD时并没有立即进行转换,仅仅是记住了数据集的逻辑操作
Ation(执行):触发Spark作业的运行,真正触发转换算子的计算
本系列主要讲解Spark中常用的函数操作:
1.RDD基本转换
本节所讲函数
基础转换操作:
(例1)
1
2
3
4
5
6
7
8
9
10
|
object Map {
def main(args: Array[String]) {
val conf =
new
SparkConf().setMaster(
"local"
).setAppName(
"map"
)
val sc =
new
SparkContext(conf)
val rdd = sc.parallelize(
1
to
10
)
//创建RDD
val map = rdd.map(_*
2
)
//对RDD中的每个元素都乘于2
map.foreach(x => print(x+
" "
))
sc.stop()
}
}
|
输出:
2 4 6 8 10 12 14 16 18 20
(RDD依赖图:红色块表示一个RDD区,黑色块表示该分区集合,下同)
(例2)
1
2
3
4
|
//...省略sc
val rdd = sc.parallelize(
1
to
5
)
val fm = rdd.flatMap(x => (
1
to x)).collect()
fm.foreach( x => print(x +
" "
))
|
输出:
1 1 2 1 2 3 1 2 3 4 1 2 3 4 5
如果是map函数其输出如下:
Range(1) Range(1, 2) Range(1, 2, 3) Range(1, 2, 3, 4) Range(1, 2, 3, 4, 5)
(RDD依赖图)
func的类型: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,当在映射的过程中不断的创建对象时就可以使用mapPartitions比map的效率要高很多,比如当向数据库写入数据时,如果使用map就需要为每个元素创建connection对象,但使用mapPartitions的话就需要为每个分区创建connetcion对象
(例3):输出有女性的名字:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
|
object MapPartitions {
//定义函数
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
_ =>
}
}
return
woman.iterator
}
def main(args: Array[String]) {
val conf =
new
SparkConf().setMaster(
"local"
).setAppName(
"mappartitions"
)
val sc =
new
SparkContext(conf)
val l = List((
"kpop"
,
"female"
),(
"zorro"
,
"male"
),(
"mobin"
,
"male"
),(
"lucy"
,
"female"
))
val rdd = sc.parallelize(l,
2
)
val mp = rdd.mapPartitions(partitionsFun)
/*val mp = rdd.mapPartitionsWithIndex(partitionsFun)*/
mp.collect.foreach(x => (print(x +
" "
)))
//将分区中的元素转换成Aarray再输出
}
}
|
输出:
kpop lucy
其实这个效果可以用一条语句完成
1
|
val mp = rdd.mapPartitions(x => x.filter(_._2 ==
"female"
)).map(x => x._1)
|
之所以不那么做是为了演示函数的定义
(RDD依赖图)
func类型:(Int, Iterator[T]) => Iterator[U]
(例4):将例3橙色的注释部分去掉即是
输出:(带了分区索引)
[0]kpop [1]lucy
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
(例5):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)
1
2
3
4
5
|
//省略
val rdd = sc.parallelize(
1
to
10
)
val sample1 = rdd.sample(
true
,
0.5
,
3
)
sample1.collect.foreach(x => print(x +
" "
))
sc.stop
|
1
2
3
4
5
6
|
//省略sc
val rdd1 = sc.parallelize(
1
to
3
)
val rdd2 = sc.parallelize(
3
to
5
)
val unionRDD = rdd1.union(rdd2)
unionRDD.collect.foreach(x => print(x +
" "
))
sc.stop
|
输出:
1 2 3 3 4 5
1
2
3
4
5
6
|
//省略sc
val rdd1 = sc.parallelize(
1
to
3
)
val rdd2 = sc.parallelize(
3
to
5
)
val unionRDD = rdd1.intersection(rdd2)
unionRDD.collect.foreach(x => print(x +
" "
))
sc.stop
|
输出:
3 4
1
2
3
4
5
|
//省略sc
val list = List(
1
,
1
,
2
,
5
,
2
,
9
,
6
,
1
)
val distinctRDD = sc.parallelize(list)
val unionRDD = distinctRDD.distinct()
unionRDD.collect.foreach(x => print(x +
" "
))
|
输出:
1 6 9 5 2
1
2
3
4
5
|
//省略
val rdd1 = sc.parallelize(
1
to
3
)
val rdd2 = sc.parallelize(
2
to
5
)
val cartesianRDD = rdd1.cartesian(rdd2)
cartesianRDD.foreach(x => println(x +
" "
))
|
输出:
(1,2) (1,3) (1,4) (1,5) (2,2) (2,3) (2,4) (2,5) (3,2) (3,3) (3,4) (3,5)
(RDD依赖图)
目,但不会报错,只是分区个数还是原来的
(例9:)
shuffle=false
1
2
3
4
|
//省略
val rdd = sc.parallelize(
1
to
16
,
4
)
val coalesceRDD = rdd.coalesce(
3
)
//当suffle的值为false时,不能增加分区数(即分区数不能从5->7)
println(
"重新分区后的分区个数:"
+coalesceRDD.partitions.size)
|
输出:
重新分区后的分区个数:3 //分区后的数据集 List(1, 2, 3, 4) List(5, 6, 7, 8) List(9, 10, 11, 12, 13, 14, 15, 16)
(例9.1:)
shuffle=true
1
2
3
4
5
|
//...省略
val rdd = sc.parallelize(
1
to
16
,
4
)
val coalesceRDD = rdd.coalesce(
7
,
true
)
println(
"重新分区后的分区个数:"
+coalesceRDD.partitions.size)
println(
"RDD依赖关系:"
+coalesceRDD.toDebugString)
|
输出:
重新分区后的分区个数:5 RDD依赖关系:(5) MapPartitionsRDD[4] at coalesce at Coalesce.scala:14 [] | CoalescedRDD[3] at coalesce at Coalesce.scala:14 [] | ShuffledRDD[2] at coalesce at Coalesce.scala:14 [] +-(4) MapPartitionsRDD[1] at coalesce at Coalesce.scala:14 [] | ParallelCollectionRDD[0] at parallelize at Coalesce.scala:13 [] //分区后的数据集 List(10, 13) List(1, 5, 11, 14) List(2, 6, 12, 15) List(3, 7, 16) List(4, 8, 9)
(RDD依赖图:coalesce(3,flase))
(RDD依赖图:coalesce(3,true))
11.repartition(numPartition):是函数coalesce(numPartition,true)的实现,效果和例9.1的coalesce(numPartition,true)的一样
1
2
3
4
5
|
//省略
val rdd = sc.parallelize(
1
to
16
,
4
)
val glomRDD = rdd.glom()
//RDD[Array[T]]
glomRDD.foreach(rdd => println(rdd.getClass.getSimpleName))
sc.stop
|
输出:
int[] //说明RDD中的元素被转换成数组Array[Int]
1
2
3
4
5
6
7
|
//省略sc
val rdd = sc.parallelize(
1
to
10
)
val randomSplitRDD = rdd.randomSplit(Array(
1.0
,
2.0
,
7.0
))
randomSplitRDD(
0
).foreach(x => print(x +
" "
))
randomSplitRDD(
1
).foreach(x => print(x +
" "
))
randomSplitRDD(
2
).foreach(x => print(x +
" "
))
sc.stop
|
输出:
2 4 3 8 9 1 5 6 7 10