Spark Core应用解析

深藏阁楼爱情的钟 2022-10-03 00:34 188阅读 0赞

# 一。RDD概念 #

## 1.1。RDD概述 ##

### 1.1.1。什么是RDD ###

RDD（Resilient Distributed Dataset）叫做分布式数据集，是Spark中最基本的数据抽象，它代表一个不可变、可分区、里面的元素可并行计算的集合。在 Spark 中，对数据的所有操作不外乎创建 RDD、转化已有RDD 以及调用 RDD 操作进行求值。每个 RDD 都被分为多个分区，这些分区运行在集群中的不同节点上。RDD 可以包含 Python、Java、Scala 中任意类型的对象， 甚至可以包含用户自定义的对象。RDD具有数据流模型的特点：自动容错、位置感知性调度和可伸缩性。RDD允许用户在执行多个查询时显式地将工作集缓存在内存中，后续的查询能够重用工作集，这极大地提升了查询速度。

RDD支持两种操作:转化操作和行动操作。RDD 的转化操作是返回一个新的 RDD的操作，比如 map()和 filter()，而行动操作则是向驱动器程序返回结果或把结果写入外部系统的操作。比如 count() 和 first()。

Spark采用惰性计算模式，RDD只有第一次在一个行动操作中用到时，才会真正计算。Spark可以优化整个计算过程。默认情况下，Spark 的 RDD 会在你每次对它们进行行动操作时重新计算。如果想在多个行动操作中重用同一个 RDD，可以使用 RDD.persist() 让 Spark 把这个 RDD 缓存下来。

### 1.1.2。RDD的属性 ###

1)   一组分片（Partition），即数据集的基本组成单位。对于RDD来说，每个分片都会被一个计算任务处理，并决定并行计算的粒度。用户可以在创建RDD时指定RDD的分片个数，如果没有指定，那么就会采用默认值。默认值就是程序所分配到的CPU Core的数目。

2)   一个计算每个分区的函数。Spark中RDD的计算是以分片为单位的，每个RDD都会实现compute函数以达到这个目的。compute函数会对迭代器进行复合，不需要保存每次计算的结果。

3)   RDD之间的依赖关系。RDD的每次转换都会生成一个新的RDD，所以RDD之间就会形成类似于流水线一样的前后依赖关系。在部分分区数据丢失时，Spark可以通过这个依赖关系重新计算丢失的分区数据，而不是对RDD的所有分区进行重新计算。

4)   一个Partitioner，即RDD的分片函数。当前Spark中实现了两种类型的分片函数，一个是基于哈希的HashPartitioner，另外一个是基于范围的RangePartitioner。只有对于于key-value的RDD，才会有Partitioner，非key-value的RDD的Parititioner的值是None。Partitioner函数不但决定了RDD本身的分片数量，也决定了parent RDD Shuffle输出时的分片数量。

5）一个列表，存储存取每个Partition的优先位置（preferred location）。对于一个HDFS文件来说，这个列表保存的就是每个Partition所在的块的位置。按照“移动数据不如移动计算”的理念，Spark在进行任务调度的时候，会尽可能地将计算任务分配到其所要处理数据块的存储位置。

RDD是一个应用层面的逻辑概念。一个RDD多个分片。RDD就是一个元数据记录集，记录了RDD内存所有的关系数据。

![1271822-20190622000805304-444959404.png][]

## 1.2。RDD弹性 ##

1)   自动进行内存和磁盘数据存储的切换

Spark优先把数据放到内存中，如果内存放不下，就会放到磁盘里面，程序进行自动的存储切换

2)   基于血统的高效容错机制

在RDD进行转换和动作的时候，会形成RDD的Lineage依赖链，当某一个RDD失效的时候，可以通过重新计算上游的RDD来重新生成丢失的RDD数据。

3)   Task如果失败会自动进行特定次数的重试

RDD的计算任务如果运行失败，会自动进行任务的重新计算，默认次数是4次。

4)   Stage如果失败会自动进行特定次数的重试

如果Job的某个Stage阶段计算失败，框架也会自动进行任务的重新计算，默认次数也是4次。

5)   Checkpoint和Persist可主动或被动触发

RDD可以通过Persist持久化将RDD缓存到内存或者磁盘，当再次用到该RDD时直接读取就行。也可以将RDD进行检查点，检查点会将数据存储在HDFS中，该RDD的所有父RDD依赖都会被移除。

6)   数据调度弹性

Spark把这个JOB执行模型抽象为通用的有向无环图DAG，可以将多Stage的任务串联或并行执行，调度引擎自动处理Stage的失败以及Task的失败。

7)   数据分片的高度弹性

可以根据业务的特征，动态调整数据分片的个数，提升整体的应用执行效率。

RDD全称叫做弹性分布式数据集(Resilient Distributed Datasets)，它是一种分布式的内存抽象，表示一个只读的记录分区的集合，它只能通过其他RDD转换而创建，为此，RDD支持丰富的转换操作(如map, join, filter, groupBy等)，通过这种转换操作，新的RDD则包含了如何从其他RDDs衍生所必需的信息，所以说RDDs之间是有依赖关系的。基于RDDs之间的依赖，RDDs会形成一个有向无环图DAG，该DAG描述了整个流式计算的流程，实际执行的时候，RDD是通过血缘关系(Lineage)一气呵成的，即使出现数据分区丢失，也可以通过血缘关系重建分区。

## 1.3。RDD特点 ##

RDD表示只读的分区的数据集，对RDD进行改动，只能通过RDD的转换操作，由一个RDD得到一个新的RDD，新的RDD包含了从其他RDD衍生所必需的信息。RDDs之间存在依赖，RDD的执行是按照血缘关系延时计算的。如果血缘关系较长，可以通过持久化RDD来切断血缘关系。

### 1.3.1 分区 ###

RDD逻辑上是分区的，每个分区的数据是抽象存在的，计算的时候会通过一个compute函数得到每个分区的数据。如果RDD是通过已有的文件系统构建，则compute函数是读取指定文件系统中的数据，如果RDD是通过其他RDD转换而来，则compute函数是执行转换逻辑将其他RDD的数据进行转换。

![1271822-20190622001511289-118180144.png][]

### 1.3.2 只读 ###

RDD是只读的，要想改变RDD中的数据，只能在现有的RDD基础上创建新的RDD。如下图所示

![1271822-20190622001616137-1729011386.png][]

由一个RDD转换到另一个RDD，可以通过丰富的操作算子实现，不再像MapReduce那样只能写map和reduce了，如下图所示。

![1271822-20190622001714777-898735087.png][]

RDD的操作算子包括两类，一类叫做transformations，它是用来将RDD进行转化，构建RDD的血缘关系；另一类叫做actions，它是用来触发RDD的计算，得到RDD的相关计算结果或者将RDD保存的文件系统中

### 1.3.3 依赖 ###

RDDs通过操作算子进行转换，转换得到的新RDD包含了从其他RDDs衍生所必需的信息，RDDs之间维护着这种血缘关系，也称之为依赖。如下图所示，依赖包括两种，一种是窄依赖，RDDs之间分区是一一对应的，另一种是宽依赖，下游RDD的每个分区与上游RDD(也称之为父RDD)的每个分区都有关，是多对多的关系。

![1271822-20190622002441466-516156718.png][]

通过RDDs之间的这种依赖关系，一个任务流可以描述为DAG(有向无环图)，如下图所示，在实际执行过程中宽依赖对应于Shuffle(图中的reduceByKey和join)，窄依赖中的所有转换操作可以通过类似于管道的方式一气呵成执行(图中map和union可以一起执行)。

![1271822-20190622002547187-635763137.png][]

### 1.3.4 缓存 ###

如果在应用程序中多次使用同一个RDD，可以将该RDD缓存起来，该RDD只有在第一次计算的时候会根据血缘关系得到分区的数据，在后续其他地方用到该RDD的时候，会直接从缓存处取而不用再根据血缘关系计算，这样就加速后期的重用。如下图所示，RDD-1经过一系列的转换后得到RDD-n并保存到hdfs，RDD-1在这一过程中会有个中间结果，如果将其缓存到内存，那么在随后的RDD-1转换到RDD-m这一过程中，就不会计算其之前的RDD-0了。

![1271822-20190622002651928-1951587523.png][]

### 1.3.5 checkpoint ###

虽然RDD的血缘关系天然地可以实现容错，当RDD的某个分区数据失败或丢失，可以通过血缘关系重建。但是对于长时间迭代型应用来说，随着迭代的进行，RDDs之间的血缘关系会越来越长，一旦在后续迭代过程中出错，则需要通过非常长的血缘关系去重建，势必影响性能。为此，RDD支持checkpoint将数据保存到持久化的存储中，这样就可以切断之前的血缘关系，因为checkpoint后的RDD不需要知道它的父RDDs了，它可以从checkpoint处拿到数据。

给定一个RDD我们至少可以知道如下几点信息：

1、分区数以及分区方式；

2、由父RDDs衍生而来的相关依赖信息；

3、计算每个分区的数据，计算步骤为：

1）如果被缓存，则从缓存中取的分区的数据；

2）如果被checkpoint，则从checkpoint处恢复数据；

3）根据血缘关系计算分区的数据。

# 二 RDD编程 #

## 2.1 编程模型 ##

在Spark中，RDD被表示为对象，通过对象上的方法调用来对RDD进行转换。经过一系列的transformations定义RDD之后，就可以调用actions触发RDD的计算，action可以是向应用程序返回结果(count, collect等)，或者是向存储系统保存数据(saveAsTextFile等)。在Spark中，只有遇到action，才会执行RDD的计算(即延迟计算)，这样在运行时可以通过管道的方式传输多个转换。

要使用Spark，需要编写一个Driver程序，它被提交到集群以调度运行Worker，如下图所示。Driver中定义了一个或多个RDD，并调用RDD上的action，Worker则执行RDD分区计算任务。

![1271822-20190622003232387-2100520195.png][]

Dirver ，SparkContext ，Executor ，Master ，Worker 关系如图

![1271822-20190622003404229-247657266.png][]

## 2.2 创建RDD ##

在Spark中创建RDD的创建方式大概可以分为三种：（1）、从集合中创建RDD；（2）、从外部存储创建RDD；（3）、从其他RDD创建。

1）由一个已经存在的Scala集合创建，集合并行化。

val rdd1 = sc.parallelize(Array(1,2,3,4,5,6,7,8))

而从集合中创建RDD，[Spark][]主要提供了两种函数：parallelize和makeRDD

makeRDD函数有两种实现，第一种实现其实完全和parallelize一致；而第二种实现可以为数据提供位置信息，而除此之外的实现和parallelize函数也是一致的。

scala> val gh01= sc.parallelize(List(1,2,3))
 gh01: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[10] at parallelize at <console>:21
 
 scala> val gh02 = sc.makeRDD(List(1,2,3))
 gh022: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[11] at makeRDD at <console>:21
 
 scala> val seq = List((1, List("Hello", "World", "Spark")),
 | (2, List("At", "zgh")))
 seq: List[(Int, List[String])] = List((1,List(Hello, World, Spark)),
 (2,List(At, zgh)))
 
 scala> val gh03 = sc.makeRDD(seq)
 gh03: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[12] at makeRDD at <console>:23
 
 scala> guigu3.preferredLocations(gh03.partitions(1))
 res26: Seq[String] = List(At, zgh)
 
 scala> gh03.preferredLocations(gh03.partitions(0))
 res27: Seq[String] = List(Hello, World, Spark)
 
 scala> gh01.preferredLocations(gh01.partitions(0))
 res28: Seq[String] = List()

2)由外部存储系统的数据集创建，包括本地的文件系统，还有所有Hadoop支持的数据集，比如HDFS、Cassandra、HBase等

scala> val atgh = sc.textFile("hdfs://master01:9000/RELEASE")
 atgh: org.apache.spark.rdd.RDD[String] = hdfs://master01:9000/RELEASE MapPartitionsRDD[4] at textFile at <console>:24

## 2.3 RDD编程 ##

RDD一般分为数值RDD和键值对RDD

### 2.3.1 Transformation ###

RDD中的所有转换都是延迟加载的，也就是说，它们并不会直接计算结果。相反的，它们只是记住这些应用到基础数据集（例如一个文件）上的转换动作。只有当发生一个要求返回结果给Driver的动作时，这些转换才会真正运行。这种设计让Spark更加有效率地运行。

**常用的Transformation：**

<table style="border-style:none;width:576px;"> 
 <tbody> 
 <tr> 
 <td style="width:183.85pt;border-top-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-top-color:#5b9bd5;border-bottom-color:#5b9bd5;border-left-color:#5b9bd5;border-right:none;background:#5b9bd5;" colspan="4"> 转换 </td> 
 <td style="width:248.05pt;border-top-width:1pt;border-right-width:1pt;border-bottom-width:1pt;border-top-color:#5b9bd5;border-right-color:#5b9bd5;border-bottom-color:#5b9bd5;border-left:none;background:#5b9bd5;"> 含义 </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> map(func) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> <img src="https://img2018.cnblogs.com/blog/1271822/201906/1271822-20190622010338695-480250653.png" alt=""> 
 <div> 
 <pre>scala&gt; var source = sc.parallelize(1 to 10)
source: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[8] at parallelize at &lt;console&gt;:24

scala&gt; source.collect()
res7: Array[Int] = Array(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

scala&gt; val mapadd = source.map(_ * 2)
mapadd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[9] at map at &lt;console&gt;:26

scala&gt; mapadd.collect()
res8: Array[Int] = Array(2, 4, 6, 8, 10, 12, 14, 16, 18, 20)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> filter(func) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; var sourceFilter = sc.parallelize(Array("xiaoming","xiaojiang","xiaohe","dazhi"))
sourceFilter: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[10] at parallelize at &lt;console&gt;:24
 
scala&gt; val filter = sourceFilter.filter(_.contains("xiao"))
filter: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[11] at filter at &lt;console&gt;:26

scala&gt; sourceFilter.collect()
res9: Array[String] = Array(xiaoming, xiaojiang, xiaohe, dazhi)

scala&gt; filter.collect()
res10: Array[String] = Array(xiaoming, xiaojiang, xiaohe)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> flatMap(func) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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） </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> &nbsp;<img src="https://img2018.cnblogs.com/blog/1271822/201906/1271822-20190622010940959-1972602153.png" alt=""> 
 <div> 
 <pre>scala&gt; val sourceFlat = sc.parallelize(1 to 5)
sourceFlat: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[12] at parallelize at &lt;console&gt;:24

scala&gt; sourceFlat.collect()
res11: Array[Int] = Array(1, 2, 3, 4, 5)

scala&gt; val flatMap = sourceFlat.flatMap(1 to _)
flatMap: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[13] at flatMap at &lt;console&gt;:26

scala&gt; flatMap.collect()
res12: Array[Int] = Array(1, 1, 2, 1, 2, 3, 1, 2, 3, 4, 1, 2, 3, 4, 5)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> mapPartitions(func) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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] =&gt; 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; :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") =&gt; woman = next._1 :: woman

case _ =&gt;

}

woman.iterator

}

// Exiting paste mode, now interpreting.

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

scala&gt; val result = rdd.mapPartitions(partitionsFun)
result: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[17] at mapPartitions at &lt;console&gt;:28

scala&gt; result.collect()
res13: Array[String] = Array(kpop, lucy)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> mapPartitionsWithIndex(func) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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]) =&gt; Iterator[U] </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; :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") =&gt; woman = "["+index+"]"+next._1 :: woman

case _ =&gt;

}

woman.iterator

}

scala&gt; val result = rdd.mapPartitionsWithIndex(partitionsFun)
result: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[19] at mapPartitionsWithIndex at &lt;console&gt;:28

scala&gt; result.collect()
res14: Array[String] = Array([0]kpop, [3]lucy)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> sample(withReplacement, fraction, seed) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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） </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(1 to 10)
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[20] at parallelize at &lt;console&gt;:24

scala&gt; rdd.collect()
res15: Array[Int] = Array(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)

scala&gt; var sample1 = rdd.sample(true,0.4,2)
sample1: org.apache.spark.rdd.RDD[Int] = PartitionwiseSampledRDD[21] at sample at &lt;console&gt;:26

scala&gt; sample1.collect()
res16: Array[Int] = Array(1, 2, 2, 7, 7, 8, 9)

scala&gt; var sample2 = rdd.sample(false,0.2,3)
sample2: org.apache.spark.rdd.RDD[Int] = PartitionwiseSampledRDD[22] at sample at &lt;console&gt;:26

scala&gt; sample2.collect()
res17: Array[Int] = Array(1, 9)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> takeSample &nbsp; </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 和Sample的区别是：takeSample返回的是最终的结果集合。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5">&nbsp;</td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> union(otherDataset) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 对源RDD和参数RDD求并集后返回一个新的RDD </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd1 = sc.parallelize(1 to 5)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[23] at parallelize at &lt;console&gt;:24

scala&gt; val rdd2 = sc.parallelize(5 to 10)
rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[24] at parallelize at &lt;console&gt;:24

scala&gt; val rdd3 = rdd1.union(rdd2)
rdd3: org.apache.spark.rdd.RDD[Int] = UnionRDD[25] at union at &lt;console&gt;:28

scala&gt; rdd3.collect()
res18: Array[Int] = Array(1, 2, 3, 4, 5, 5, 6, 7, 8, 9, 10)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> intersection(otherDataset) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 对源RDD和参数RDD求交集后返回一个新的RDD </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd1 = sc.parallelize(1 to 7)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[26] at parallelize at &lt;console&gt;:24

scala&gt; val rdd2 = sc.parallelize(5 to 10)
rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[27] at parallelize at &lt;console&gt;:24

scala&gt; val rdd3 = rdd1.intersection(rdd2)
rdd3: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[33] at intersection at &lt;console&gt;:28

scala&gt; rdd3.collect()
[Stage 15:=============================&gt; (2 + 2) res19: Array[Int] = Array(5, 6, 7)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> distinct([numTasks])) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; val unionRDD = distinctRdd.distinct()
unionRDD: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[37] at distinct at &lt;console&gt;:26

scala&gt; unionRDD.collect()
[Stage 16:&gt; (0 + 4) [Stage 16:=============================&gt; (2 + 2) res20: Array[Int] = Array(1, 9, 5, 6, 2)

scala&gt; val unionRDD = distinctRdd.distinct(2)
unionRDD: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[40] at distinct at &lt;console&gt;:26

scala&gt; unionRDD.collect()
res21: Array[Int] = Array(6, 2, 1, 9, 5)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:162.6pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="2"> partitionBy &nbsp; </td> 
 <td style="width:269.3pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;" colspan="3"> 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，&nbsp; 否则会生成ShuffleRDD.&nbsp; </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; rdd.partitions.size
res24: Int = 4

scala&gt; var rdd2 = rdd.partitionBy(new org.apache.spark.HashPartitioner(2))
rdd2: org.apache.spark.rdd.RDD[(Int, String)] = ShuffledRDD[45] at partitionBy at &lt;console&gt;:26

scala&gt; rdd2.partitions.size
res25: Int = 2</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> reduceByKey(func, [numTasks]) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> &nbsp;<img src="https://img2018.cnblogs.com/blog/1271822/201906/1271822-20190622012429638-1013120643.png" alt=""> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; val reduce = rdd.reduceByKey((x,y) =&gt; x+y)
reduce: org.apache.spark.rdd.RDD[(String, Int)] = ShuffledRDD[47] at reduceByKey at &lt;console&gt;:26

scala&gt; reduce.collect()
res29: Array[(String, Int)] = Array((female,6), (male,7))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:148.4pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;"> groupByKey </td> 
 <td style="width:10cm;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;" colspan="4"> groupByKey也是对每个key进行操作，但只生成一个sequence。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> &nbsp;<img src="https://img2018.cnblogs.com/blog/1271822/201906/1271822-20190622012511810-1875202517.png" alt=""> 
 <div> 
 <pre>scala&gt; val words = Array("one", "two", "two", "three", "three", "three")
words: Array[String] = Array(one, two, two, three, three, three)

scala&gt; val wordPairsRDD = sc.parallelize(words).map(word =&gt; (word, 1))
wordPairsRDD: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[4] at map at &lt;console&gt;:26

scala&gt; val group = wordPairsRDD.groupByKey()
group: org.apache.spark.rdd.RDD[(String, Iterable[Int])] = ShuffledRDD[5] at groupByKey at &lt;console&gt;:28

scala&gt; group.collect()
res1: Array[(String, Iterable[Int])] = Array((two,CompactBuffer(1, 1)), (one,CompactBuffer(1)), (three,CompactBuffer(1, 1, 1)))

scala&gt; group.map(t =&gt; (t._1, t._2.sum))
res2: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[6] at map at &lt;console&gt;:31

scala&gt; res2.collect()
res3: Array[(String, Int)] = Array((two,2), (one,1), (three,3))

scala&gt; val map = group.map(t =&gt; (t._1, t._2.sum))
map: org.apache.spark.rdd.RDD[(String, Int)] = MapPartitionsRDD[7] at map at &lt;console&gt;:30

scala&gt; map.collect()
res4: Array[(String, Int)] = Array((two,2), (one,1), (three,3))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:148.4pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;"> combineByKey[C](&nbsp;&nbsp; createCombiner:&nbsp;V&nbsp;=&gt;&nbsp;C,&nbsp;&nbsp; mergeValue:&nbsp;(C,&nbsp;V)&nbsp;=&gt;&nbsp;C,&nbsp;&nbsp; mergeCombiners:&nbsp;(C,&nbsp;C)&nbsp;=&gt;&nbsp;C)&nbsp; &nbsp; </td> 
 <td style="width:10cm;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;" colspan="4"> 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&nbsp;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&nbsp; 那个键对应的累加器的初始值 mergeValue:&nbsp;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:&nbsp;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。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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&gt; val input = sc.parallelize(scores)
input: org.apache.spark.rdd.RDD[(String, Int)] = ParallelCollectionRDD[52] at parallelize at &lt;console&gt;:26

scala&gt; val combine = input.combineByKey(

| (v)=&gt;(v,1),

| (acc:(Int,Int),v)=&gt;(acc._1+v,acc._2+1),

| (acc1:(Int,Int),acc2:(Int,Int))=&gt;(acc1._1+acc2._1,acc1._2+acc2._2))

combine: org.apache.spark.rdd.RDD[(String, (Int, Int))] = ShuffledRDD[53] at combineByKey at &lt;console&gt;:28

scala&gt; val result = combine.map{

| case (key,value) =&gt; (key,value._1/value._2.toDouble)}

result: org.apache.spark.rdd.RDD[(String, Double)] = MapPartitionsRDD[54] at map at &lt;console&gt;:30

scala&gt; result.collect()
res33: Array[(String, Double)] = Array((Wilma,95.33333333333333), (Fred,91.33333333333333))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> aggregateByKey(zeroValue:U,[partitioner:&nbsp;Partitioner])(seqOp: (U, V) =&gt; U,combOp: (U, U) =&gt; U) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 例如：分一个分区，以key为1的分区为例，0先和3比较得3，3在和2比较得3，3在和4比较得4，所以整个key为1的组最终结果为（1，4），同理，key为2的最终结果为（2，3），key为3的为（3，8）. 如果分三个分区，前两个是一个分区，中间两个是一个分区，最后两个是一个分区，第一个分区的最终结果为（1，3），第二个分区为（1，4）（2，3），最后一个分区为（3，8），combine后为 (3,8), (1,7), (2,3) &nbsp;<img src="https://img2018.cnblogs.com/blog/1271822/201906/1271822-20190622012756601-404305171.png" alt=""> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; val agg = rdd.aggregateByKey(0)(math.max(_,_),_+_)
agg: org.apache.spark.rdd.RDD[(Int, Int)] = ShuffledRDD[13] at aggregateByKey at &lt;console&gt;:26

scala&gt; agg.collect()
res7: Array[(Int, Int)] = Array((3,8), (1,7), (2,3))

scala&gt; agg.partitions.size
res8: Int = 3

scala&gt; 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 &lt;console&gt;:24

scala&gt; val agg = rdd.aggregateByKey(0)(math.max(_,_),_+_).collect()
agg: Array[(Int, Int)] = Array((1,4), (3,8), (2,3))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> foldByKey(zeroValue: V)(func: (V, V) =&gt; V): RDD[(K, V)] </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> aggregateByKey的简化操作，seqop和combop相同 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; val agg = rdd.foldByKey(0)(_+_)
agg: org.apache.spark.rdd.RDD[(Int, Int)] = ShuffledRDD[92] at foldByKey at &lt;console&gt;:26

scala&gt; agg.collect()
res61: Array[(Int, Int)] = Array((3,14), (1,9), (2,3))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> sortByKey([ascending], [numTasks]) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> &nbsp; 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; rdd.sortByKey(true).collect()
res9: Array[(Int, String)] = Array((1,dd), (2,bb), (3,aa), (6,cc))

scala&gt; rdd.sortByKey(false).collect()
res10: Array[(Int, String)] = Array((6,cc), (3,aa), (2,bb), (1,dd))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> sortBy(func,[ascending], [numTasks]) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 与sortByKey类似，但是更灵活,可以用func先对数据进行处理，按照处理后的数据比较结果排序。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(List(1,2,3,4))
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[21] at parallelize at &lt;console&gt;:24

scala&gt; rdd.sortBy(x =&gt; x).collect()
res11: Array[Int] = Array(1, 2, 3, 4)

scala&gt; rdd.sortBy(x =&gt; x%3).collect()
res12: Array[Int] = Array(3, 4, 1, 2)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> join(otherDataset, [numTasks]) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(Array((1,"a"),(2,"b"),(3,"c")))
rdd: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[32] at parallelize at &lt;console&gt;:24

scala&gt; val rdd1 = sc.parallelize(Array((1,4),(2,5),(3,6)))
rdd1: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[33] at parallelize at &lt;console&gt;:24

scala&gt; rdd.join(rdd1).collect()
res13: Array[(Int, (String, Int))] = Array((1,(a,4)), (2,(b,5)), (3,(c,6)))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> cogroup(otherDataset, [numTasks]) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 在类型为(K,V)和(K,W)的RDD上调用，返回一个(K,(Iterable&lt;V&gt;,Iterable&lt;W&gt;))类型的RDD </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> &nbsp;<img src="https://img2018.cnblogs.com/blog/1271822/201906/1271822-20190622013127771-1582995866.png" alt=""> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(Array((1,"a"),(2,"b"),(3,"c")))
rdd: org.apache.spark.rdd.RDD[(Int, String)] = ParallelCollectionRDD[37] at parallelize at &lt;console&gt;:24

scala&gt; val rdd1 = sc.parallelize(Array((1,4),(2,5),(3,6)))
rdd1: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[38] at parallelize at &lt;console&gt;:24

scala&gt; 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&gt; val rdd2 = sc.parallelize(Array((4,4),(2,5),(3,6)))
rdd2: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[41] at parallelize at &lt;console&gt;:24

scala&gt; 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&gt; 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 &lt;console&gt;:24

scala&gt; rdd3.cogroup(rdd2).collect()
[Stage 36:&gt; (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))))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> cartesian(otherDataset) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 笛卡尔积 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd1 = sc.parallelize(1 to 3)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[47] at parallelize at &lt;console&gt;:24

scala&gt; val rdd2 = sc.parallelize(2 to 5)
rdd2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[48] at parallelize at &lt;console&gt;:24

scala&gt; 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))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> pipe(command, [envVars]) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 对于每个分区，都执行一个perl或者shell脚本，返回输出的RDD </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(List("hi","Hello","how","are","you"),1)
rdd: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[50] at parallelize at &lt;console&gt;:24

scala&gt; rdd.pipe("/home/bigdata/pipe.sh").collect()
res18: Array[String] = Array(AA, &gt;&gt;&gt;hi, &gt;&gt;&gt;Hello, &gt;&gt;&gt;how, &gt;&gt;&gt;are, &gt;&gt;&gt;you)

scala&gt; val rdd = sc.parallelize(List("hi","Hello","how","are","you"),2)
rdd: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[52] at parallelize at &lt;console&gt;:24

scala&gt; rdd.pipe("/home/bigdata/pipe.sh").collect()
res19: Array[String] = Array(AA, &gt;&gt;&gt;hi, &gt;&gt;&gt;Hello, AA, &gt;&gt;&gt;how, &gt;&gt;&gt;are, &gt;&gt;&gt;you)

pipe.sh:

#!/bin/sh

echo "AA"

while read LINE; do

echo "&gt;&gt;&gt;"${LINE}

done</pre> 
 </div> 注意：shell脚本需要集群中的所有节点都能访问到。 </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> coalesce(numPartitions)&nbsp;&nbsp;&nbsp;&nbsp; </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(1 to 16,4)
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[54] at parallelize at &lt;console&gt;:24

scala&gt; rdd.partitions.size
res20: Int = 4

scala&gt; val coalesceRDD = rdd.coalesce(3)
coalesceRDD: org.apache.spark.rdd.RDD[Int] = CoalescedRDD[55] at coalesce at &lt;console&gt;:26

scala&gt; coalesceRDD.partitions.size
res21: Int = 3</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> repartition(numPartitions) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 根据分区数，从新通过网络随机洗牌所有数据。 </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(1 to 16,4)
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[56] at parallelize at &lt;console&gt;:24

scala&gt; rdd.partitions.size
res22: Int = 4

scala&gt; val rerdd = rdd.repartition(2)
rerdd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[60] at repartition at &lt;console&gt;:26

scala&gt; rerdd.partitions.size
res23: Int = 2

scala&gt; val rerdd = rdd.repartition(4)
rerdd: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[64] at repartition at &lt;console&gt;:26

scala&gt; rerdd.partitions.size
res24: Int = 4</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> repartitionAndSortWithinPartitions(partitioner) </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 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。&nbsp; </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="4"> &nbsp; </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;"> &nbsp; </td> 
 </tr> 
 <tr> 
 <td style="width:183.85pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="4"> glom </td> 
 <td style="width:248.05pt;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;"> 将每一个分区形成一个数组，形成新的RDD类型时RDD[Array[T]] </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(1 to 16,4)
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[65] at parallelize at &lt;console&gt;:24

scala&gt; 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))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:176.75pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="3"> mapValues </td> 
 <td style="width:9cm;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;" colspan="2"> 针对于(K,V)形式的类型只对V进行操作&nbsp; </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; 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 &lt;console&gt;:24

scala&gt; rdd3.mapValues(_+"|||").collect()
res26: Array[(Int, String)] = Array((1,a|||), (1,d|||), (2,b|||), (3,c|||))</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:176.75pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="3"> subtract </td> 
 <td style="width:9cm;border-top:none;border-left:none;border-bottom-width:1pt;border-bottom-color:#9cc2e5;border-right-width:1pt;border-right-color:#9cc2e5;background:#deeaf6;" colspan="2"> 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&nbsp; </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;" colspan="5"> 
 <div> 
 <pre>scala&gt; val rdd = sc.parallelize(3 to 8)
rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[70] at parallelize at &lt;console&gt;:24

scala&gt; val rdd1 = sc.parallelize(1 to 5)
rdd1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[71] at parallelize at &lt;console&gt;:24

scala&gt; rdd.subtract(rdd1).collect()
res27: Array[Int] = Array(8, 6, 7)</pre> 
 </div> </td> 
 </tr> 
 <tr> 
 <td style="width:431.9pt;border-right-width:1pt;border-bottom-width:1pt;border-left-width:1pt;border-right-color:#9cc2e5;border-bottom-color:#9cc2e5;border-left-color:#9cc2e5;border-top:none;background:#deeaf6;" colspan="5"> &nbsp; </td> 
 </tr> 
 </tbody> 
</table>

### 2.3.2 Action ###

**常用的Action：**

转载于:https://www.cnblogs.com/GH-123/p/11067551.html

[1271822-20190622000805304-444959404.png]: /images/20220109/65c13e50afe641ea97b86ba45f38f93d.png
[1271822-20190622001511289-118180144.png]: /images/20220109/eb4bd8e0dbbb4ea49e414eae8b008974.png
[1271822-20190622001616137-1729011386.png]: /images/20220109/5a1b95c6e90e480987507e08cd11ed66.png
[1271822-20190622001714777-898735087.png]: /images/20220109/94874c8c9cee4dbf97924dcec9951964.png
[1271822-20190622002441466-516156718.png]: /images/20220109/9b75ec82bb8743cea21d6e323f68f841.png
[1271822-20190622002547187-635763137.png]: /images/20220109/17fc641baf9544e5ba8c2f94fe9301e7.png
[1271822-20190622002651928-1951587523.png]: /images/20220109/2a94bfea56fa41d4822e52798ab76f17.png
[1271822-20190622003232387-2100520195.png]: /images/20220109/2e8b1e82369b40b59d733e96e5770da7.png
[1271822-20190622003404229-247657266.png]: /images/20220109/30a8ec80cb9243e4bdc6f5e01b79d5f9.png
[Spark]: https://www.iteblog.com/archives/tag/spark/