Spark RDD编程（Python和Scala版本）

Spark中的RDD就是一个不可变的分布式对象集合，是一种具有兼容性的基于内存的集群计算抽象方法，Spark则是这个方法的抽象。

Spark的RDD操作分为转化操作（transformation）和行动操作（action），两者的区别在于：

       a.转化操作返回一个新的RDD对象

       b.行动操作则会对RDD产生一个计算结果，并把结果返回到驱动器程序中或者把结果存储到外部存储系统（如HDFS）

常见的转化操作有：map，filter，flatMap，sample，union，distinct，

                                    groupByKey，reduceByKey，sortByKey，join，cogroup，cartesian  ......

常见的行动操作有：reduce，collect，count，first，take，taksSample，

                                    saveAsTextFile，saveAsSequenceFile，countByKey，foreach ......

下面我们以实例说明Saprk的RDD编程

1：创建RDD

      有两种方式：读取外部数据集，以及在驱动器程序中对一个集合进行并行化

       python：

>>> nums =sc.parallelize([1,2,3,4])>>> numsParallelCollectionRDD[0] at parallelize at PythonRDD.scala:423

>>> words = sc.textFile("file:///usr/local/hadoop/spark/README.md")>>> wordsfile:///usr/local/hadoop/spark/README.md MapPartitionsRDD[2] at textFile at NativeMethodAccessorImpl.java:-2>>> 

       Scala(两种方法)：

val lines = sc.parallelize(List(1,2,3,4))lines: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[0] at parallelize at <console>:27

val rdd = sc.makeRDD(1 to 10,2)rdd: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[10] at makeRDD at <console>:27

2：map()函数 和 take()函数

      RDD.map(func)，map接受一个函数作为参数，作用于RDD中的每个对象，并将返回结果作为结果RDD中对应的元素的值

      RDD.take(num)，用于取回num个value，在这里结合map使用，方便查看值

Python：

>>> nums = sc.parallelize([1,2,3,4])>>> for num in nums.take(4):...     print num... 1234>>> new_nums = nums.map(lambda x: x*2)>>> for new_num in new_nums.take(4):...     print new_num... 2468

Scala：

scala> val nums = sc.parallelize(List(1,2,3,4))nums: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[4] at parallelize at <console>:27scala>nums.take(4).foreach(println)1234

3：flatMap()函数

      RDD.flatMap(func)，和map类似，只不过map返回的是一个个元素，而flatMap返回的则是一个返回值序列的迭代器

Python：

>>> string = sc.parallelize(["i love you"])>>> new_str = string.flatMap(lambda str:str.split(" "))</span>>>> for str in new_str.take(3):...     print str... iloveyou

Scala：

scala> val string = sc.parallelize(List("i love you"))string: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[5] at parallelize at <console>:27scala> val new_str = string.flatMap(line=>line.split(" "))new_str: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[6] at flatMap at <console>:29scala> new_str.take(3).foreach(println)iloveyou

4：filter()函数和 first()函数

      RDD.filter(func)，接受一个函数作为参数，并将RDD中满足该函数的元素放入新的RDD中返回

      RDD.first()，返回的第一个

Python：

>>> string = sc.parallelize(["i love you"])>>> new_str = string.filter(lambda line : "you" in line)>>> new_str.first()'i love you'

Scala：

scala> val string = sc.parallelize(List("i love you"))string: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[7] at parallelize at <console>:27scala> string.first()res3: String = i love youscala> <pre name="code" class="java">scala> val string = sc.parallelize(List("I love you"))scala> val new_str = string.filter(line =>line.contains("love"))new_str: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[6] at filter at <console>:23scala> new_str.foreach(println)I love you

5：union()函数

      RDD1.union(RDD2)，操作对象为两个RDD，返回一个新的RDD，转化操作可以操作任意数量的输入RDD

Python：

>>> num1 = sc.parallelize([1,2,3])>>> num2 = sc.parallelize([4,5,6])>>> num3 = num1.union(num2)>>> for num in num3.take(6):...     print num... 123456

Scala:

scala> val num1 = sc.parallelize(List(1,2,3))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[0] at parallelize at <console>:27scala> val num2 = sc.parallelize(List(4,5,6))num2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[1] at parallelize at <console>:27scala> val num3 = num1.union(num2)mum3: org.apache.spark.rdd.RDD[Int] = UnionRDD[2] at union at <console>:31scala> num3.count()res1: Long = 6                                                                  scala> num3.foreach(println)312456

6：count()函数和collect()函数

      RDD.count()，是统计RDD中元素的个数，返回的是一个整数

      EDD.collect()，用来收集数据，保存在一个新的数据结构中，用来持久化，需要注意的是collect不能用在大规模数据集上

Python：

>>> nums = sc.parallelize([1,2,3,4])>>> nums.count()[Stage 0:>                                                          (0 +[Stage 0:>                                                          (0 +[Stage 0:==============>   (1 + 4       >>> 

>>> new_nums = nums.collect()>>> new_nums[1, 2, 3, 4]>>> 

Scala：

scala> val num1 = sc.parallelize(List(1,2,3))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[3] at parallelize at <console>:27scala> num1.count()res3: Long = 3scala> val num2=num1.collect()num2: Array[Int] = Array(1, 2, 3)scala> num2res4: Array[Int] = Array(1, 2, 3)scala> 

7：伪集合操作

(1)：RDD.distinct，去重，但其操作的开销大，因为它需要所有数据通过网络进行混洗

Python:

>>> nums1 = sc.parallelize([1,2,3,3])>>> nums1.count()4>>> nums2=nums1.distinct()>>> nums2.count()3>>>

Scala:

scala> val num1 = sc.parallelize(List(1,2,3,3))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[4] at parallelize at <console>:27scala> val num2 = num1.distinct()num2: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[7] at distinct at <console>:29scala> num2.foreach(println)231

(2)：RDD1.intersection(RDD2),返回两个RDD中都有的元素，类似于集合中的交集

Python:

>>> nums_1=sc.parallelize([1,2,3,4,5])>>> nums_2=sc.parallelize([3,4,5,6,7])>>> nums_3=nums_1.intersection(nums_2)>>> nums_3.count()[Stage 7:>                                                          (0 +                                                                        3       >>> for num in nums_3.take(3):...     print num...345>>>

Scala:

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[9] at parallelize at <console>:27scala> val num2 = sc.parallelize(List(3,4,5,6))num2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[10] at parallelize at <console>:27scala> val num3 = num1.intersection(num2)num3: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[16] at intersection at <console>:31scala> num3.foreach(println)43

(3)：RDD1.subtract(RDD2),接受一个RDD作为参数，返回一个由只存在第一个RDD1而不存在与第二个RDD2中的所有元素组成的RDD

Python:

>>> nums_4 = nums_1.subtract(nums_2)>>> nums_4.count()2>>> for num in nums_4.take(2):...     print num...12>>>

Scala:

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[17] at parallelize at <console>:27scala> val num2 = sc.parallelize(List(3,4,5,6))num2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[18] at parallelize at <console>:27scala> val num3 = num1.subtract(num2)num3: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[22] at subtract at <console>:31scala> num3.foreach(println)21

(4)：RDD1.cartesian(RDD2),求笛卡尔积，求出所有可能的(a,b)对

Python:

>>> nums_5 = nums_1.cartesian(nums_2)>>> nums_5org.apache.spark.api.java.JavaPairRDD@5617ade8>>> nums_5.first()(1, 3)>>>

Scala：

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[23] at parallelize at <console>:27scala> val num2 = sc.parallelize(List(3,4,5,6))num2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[24] at parallelize at <console>:27scala> val num3 = num1.cartesian(num2)num3: org.apache.spark.rdd.RDD[(Int, Int)] = CartesianRDD[25] at cartesian at <console>:31scala> num3.foreach(println)(1,3)(1,5)(1,6)(1,4)(2,3)(2,4)(3,3)(2,5)(2,6)(3,4)(3,6)(4,3)(3,5)(4,5)(4,4)(4,6)

8：reduce()函数

  RDD.reduce(func),接受一个函数作为参数，操作两个RDD的元素类型的数据并返回一个同样类型的新元素

Python:

>>> nums=sc.parallelize([1,2,3,4,5,6])>>> nums.reduce(lambda x,y:x+y)21>>> 

Scala：

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[27] at parallelize at <console>:27scala> val num2 = num1.reduce((x,y)=>x+y)num2: Int = 10

9：aggregate()函数

aggregate()函数需要我们提供期待返回的类型的初始值，然后通过一个函数把RDD中的元素合并起来放入累加器，考虑到每个节点是在本地累加的，最终，还需要通过第二个函数把累加器兩兩合并

Python：

>>> nums = sc.parallelize([1,2,3,4])>>> sumCount = nums.aggregate( (0,0),... (lambda acc,value:(acc[0]+value,acc[1]+1)),... (lambda acc1,acc2:(acc1[0]+acc2[0],acc1[1]+acc2[1])))>>> sumCount[0]/float(sumCount[1])2.5>>>

Scala：

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[30] at parallelize at <console>:27scala> val result = num1.aggregate((0,0))(     | (acc,value) => (acc._1 + value,acc._2+1),     | (acc1,acc2) =>(acc1._1+acc2._1,acc1._2+acc2._2)     | )result: (Int, Int) = (10,4)scala> val avg = result._1/result._2.toDoubleavg: Double = 2.5scala> 

10：top()函数和 foreach()函数

        RDD.top(num)，从RDD中返回前边的num个元素

Python：

>>> nums = sc.parallelize([1,2,3,4])>>> new_nums = nums.top(3)>>> new_nums[4, 3, 2]

Scala：

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[31] at parallelize at <console>:27scala> num1.top(2)res10: Array[Int] = Array(4, 3)scala> 

        RDD.foreach(func)，对RDD中的每个元素使用给定的函数

Python：

>>> nums = sc.parallelize([1,2,3])>>> def add(x):...     print "\n","x+2:",x+2... >>> nums.foreach(add)x+2: 5x+2: 3x+2: 4

Scala：

scala> def add(x:Int)={     |  println (x+2)     | }add: (x: Int)Unitscala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[36] at parallelize at <console>:27scala> num1.foreach(add)6534

11：sample()函数 和 takeSample()函数

            sample(withReplacement,traction,[send])：对RDD采样以及是否转换

Python：

>>> nums = sc.parallelize([1,2,3,4,5,6,7])>>> new_nums = nums.sample(False,0.5)>>> new_numsPythonRDD[106] at RDD at PythonRDD.scala:43>>> new_nums.count()5>>> for n in new_nums.take(5):...     print n... 13567

Scala:

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[37] at parallelize at <console>:27scala> val num2 = num1.sample(false,0.5)num2: org.apache.spark.rdd.RDD[Int] = PartitionwiseSampledRDD[38] at sample at <console>:29scala> num2.foreach(println)23

          RDD.takeSample(withReplacement,num,[send]),从RDD中返回任意一些元素

Python：

>>> nums = sc.parallelize([1,2,3,4,5,6,7])>>> new_nums= nums.takeSample(False,5)>>> new_nums[5, 3, 4, 6, 7]

Scala：

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[39] at parallelize at <console>:27scala> val num2 = num1.takeSample(false,2)num2: Array[Int] = Array(3, 4)

12：persist  和 unpersist

        RDD.persist()，不带参数默认把数据以序列化的形式缓存在JVM的堆空间中

        RDD.unpersist()，手动把持久化的RDD从内存中删除

>>> nums = sc.parallelize([1,2,3,4,5,6,7])>>> new_nums = nums.persist()>>> new_numsParallelCollectionRDD[124] at parallelize at PythonRDD.scala:423>>> new_nums.unpersist()ParallelCollectionRDD[124] at parallelize at PythonRDD.scala:423>>> 

Scala：

scala> val num1 = sc.parallelize(List(1,2,3,4))num1: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[41] at parallelize at <console>:27scala> val num2 = num1.persist()num2: num1.type = ParallelCollectionRDD[41] at parallelize at <console>:27scala> num2.foreach(println)3124scala> num2.unpersist()res17: num2.type = ParallelCollectionRDD[41] at parallelize at <console>:27