RDD之四:Value型Transformation算子

处理数据类型为Value型的Transformation算子可以根据RDD变换算子的输入分区与输出分区关系分为以下几种类型:

1)输入分区与输出分区一对一型 
2)输入分区与输出分区多对一型 
3)输入分区与输出分区多对多型 
4)输出分区为输入分区子集型 
5)还有一种特殊的输入与输出分区一对一的算子类型:Cache型。 Cache算子对RDD分区进行缓存

输入分区与输出分区一对一型

(1)map

1.map(func):数据集中的每个元素经过用户自定义的函数转换形成一个新的RDD,新的RDD叫MappedRDD.

package com.sf.transform.base;

import org.apache.spark.SparkContext
import org.apache.spark.SparkConf

    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)

(2)flatMap

与map类似,但每个元素输入项都可以被映射到0个或多个的输出项,最终将结果”扁平化“后输出。

package com.sf.transform.base

import org.apache.spark.SparkContext
import org.apache.spark.SparkConf

object FlatMap {
  def main(args: Array[String]) {
    val conf = new SparkConf().setMaster("local").setAppName("map")
    val sc = new SparkContext(conf)
    val rdd = sc.parallelize(1 to 5)
    val fm = rdd.map(x => (1 to x)).collect()
    fm.foreach(x => print(x + " "))
    sc.stop()
  }
}
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)

 

(3)mapPartitions

mapPartitions函数获取到每个分区的迭代器,在函数中通过这个分区整体的迭代器对整个分区的元素进行操作。 内部实现是生成MapPartitionsRDD。

类似与map,map作用于每个分区的每个元素,但mapPartitions作用于每个分区中

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:
package com.sf.transform.base

import org.apache.spark.SparkContext
import org.apache.spark.SparkConf

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 list = List(("kpop", "female"), ("zorro", "male"), ("mobin", "male"), ("lucy", "female"))
    val rdd = sc.parallelize(list, 2) //第二个参数是分区数
    
    val mp = rdd.mapPartitions(partitionsFun)
    /*val mp = rdd.mapPartitionsWithIndex(partitionsFun)*/
    mp.collect.foreach(x => (println(x + " "))) //将分区中的元素转换成Array再输出
  }
}
输出:
kpop lucy
其实这个效果可以用一条语句完成
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)

 4.mapPartitionsWithIndex(func):与mapPartitions类似,不同的是函数多了个分区索引的参数

func类型:(Int, Iterator[T]) => Iterator[U]
(例4):将例3橙色的注释部分去掉即是
package com.sf.transform.base

import org.apache.spark.SparkContext
import org.apache.spark.SparkConf

object MapPartitions {
  //定义函数
  def partitionsFun( iter: Iterator[(String, String)]): Iterator[String] = {
    var woman = List[String]()
    while (iter.hasNext) {
      val next = iter.next()
      next match {
        case (_, "female") => woman = next._1 :: woman
        case _ =>
      }
    }
    return woman.iterator
  }
  
  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 list = List(("kpop", "female"), ("zorro", "male"), ("mobin", "male"), ("lucy", "female"))
    val rdd = sc.parallelize(list, 2) //第二个参数是分区数
    
    val mp = rdd.mapPartitions(partitionsFun)
    println()
    mp.collect.foreach(x => (print(x + " "))) //将分区中的元素转换成Array再输出
    println()
    val mp2 = rdd.mapPartitionsWithIndex(partitionsFun)
    mp2.collect.foreach(x => (print(x + " "))) //将分区中的元素转换成Array再输出
    
  }
}

输出:(带了分区索引)

[Stage 0:>                                                          (0 + 0) / 2]
                                                                                kpop lucy 
[0]kpop [1]lucy

(4)glom

glom函数将每个分区形成一个数组,内部实现是返回的RDD[Array[T]]。 

图中,方框代表一个分区。 该图表示含有V1、 V2、 V3的分区通过函数glom形成一个数组Array[(V1),(V2),(V3)]。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return an RDD created by coalescing all elements within each partition into an array.  
  3.  */  
  4. def glom(): RDD[Array[T]] = {  
  5.   new MapPartitionsRDD[Array[T], T](this, (context, pid, iter) => Iterator(iter.toArray))  
  6. }  

输入分区与输出分区多对一型

(1)union

使用union函数时需要保证两个RDD元素的数据类型相同,返回的RDD数据类型和被合并的RDD元素数据类型相同,并不进行去重操作,保存所有元素。如果想去重,可以使用distinct()。++符号相当于uion函数操作。

图中,左侧的大方框代表两个RDD,大方框内的小方框代表RDD的分区。右侧大方框代表合并后的RDD,大方框内的小方框代表分区。含有V1,V2…U4的RDD和含有V1,V8…U8的RDD合并所有元素形成一个RDD。V1、V1、V2、V8形成一个分区,其他元素同理进行合并。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return the union of this RDD and another one. Any identical elements will appear multiple  
  3.  * times (use `.distinct()` to eliminate them).  
  4.  */  
  5. def union(other: RDD[T]): RDD[T] = {  
  6.   if (partitioner.isDefined && other.partitioner == partitioner) {  
  7.     new PartitionerAwareUnionRDD(sc, Array(this, other))  
  8.   } else {  
  9.     new UnionRDD(sc, Array(this, other))  
  10.   }  
  11. }  

(2)certesian

对两个RDD内的所有元素进行笛卡尔积操作。操作后,内部实现返回CartesianRDD。 

左侧的大方框代表两个RDD,大方框内的小方框代表RDD的分区。右侧大方框代表合并后的RDD,大方框内的小方框代表分区。 
大方框代表RDD,大方框中的小方框代表RDD分区。 例如,V1和另一个RDD中的W1、 W2、 Q5进行笛卡尔积运算形成(V1,W1)、(V1,W2)、(V1,Q5)。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return the Cartesian product of this RDD and another one, that is, the RDD of all pairs of  
  3.  * elements (a, b) where a is in `this` and b is in `other`.  
  4.  */  
  5. def cartesian[U: ClassTag](other: RDD[U]): RDD[(T, U)] = new CartesianRDD(sc, this, other)  

输入分区与输出分区多对多型

groupBy

将元素通过函数生成相应的Key,数据就转化为Key-Value格式,之后将Key相同的元素分为一组。

val cleanF = sc.clean(f)中sc.clean函数将用户函数预处理; 
this.map(t => (cleanF(t), t)).groupByKey(p)对数据map进行函数操作,再对groupByKey进行分组操作。其中,p中确定了分区个数和分区函数,也就决定了并行化的程度。


图中,方框代表一个RDD分区,相同key的元素合并到一个组。 例如,V1,V2合并为一个Key-Value对,其中key为“ V” ,Value为“ V1,V2” ,形成V,Seq(V1,V2)。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return an RDD of grouped items. Each group consists of a key and a sequence of elements  
  3.  * mapping to that key. The ordering of elements within each group is not guaranteed, and  
  4.  * may even differ each time the resulting RDD is evaluated.  
  5.  *  
  6.  * Note: This operation may be very expensive. If you are grouping in order to perform an  
  7.  * aggregation (such as a sum or average) over each key, using [[PairRDDFunctions.aggregateByKey]]  
  8.  * or [[PairRDDFunctions.reduceByKey]] will provide much better performance.  
  9.  */  
  10. def groupBy[K](f: T => K)(implicit kt: ClassTag[K]): RDD[(K, Iterable[T])] =  
  11.   groupBy[K](f, defaultPartitioner(this))  
  12.   
  13. /**  
  14.  * Return an RDD of grouped elements. Each group consists of a key and a sequence of elements  
  15.  * mapping to that key. The ordering of elements within each group is not guaranteed, and  
  16.  * may even differ each time the resulting RDD is evaluated.  
  17.  *  
  18.  * Note: This operation may be very expensive. If you are grouping in order to perform an  
  19.  * aggregation (such as a sum or average) over each key, using [[PairRDDFunctions.aggregateByKey]]  
  20.  * or [[PairRDDFunctions.reduceByKey]] will provide much better performance.  
  21.  */  
  22. def groupBy[K](f: T => K, numPartitions: Int)(implicit kt: ClassTag[K]): RDD[(K, Iterable[T])] =  
  23.   groupBy(f, new HashPartitioner(numPartitions))  
  24.   
  25. /**  
  26.  * Return an RDD of grouped items. Each group consists of a key and a sequence of elements  
  27.  * mapping to that key. The ordering of elements within each group is not guaranteed, and  
  28.  * may even differ each time the resulting RDD is evaluated.  
  29.  *  
  30.  * Note: This operation may be very expensive. If you are grouping in order to perform an  
  31.  * aggregation (such as a sum or average) over each key, using [[PairRDDFunctions.aggregateByKey]]  
  32.  * or [[PairRDDFunctions.reduceByKey]] will provide much better performance.  
  33.  */  
  34. def groupBy[K](f: T => K, p: Partitioner)(implicit kt: ClassTag[K], ord: Ordering[K] = null)  
  35.     : RDD[(K, Iterable[T])] = {  
  36.   val cleanF = sc.clean(f)  
  37.   this.map(t => (cleanF(t), t)).groupByKey(p)  
  38. }  

输出分区为输入分区子集型

(1)filter

filter的功能是对元素进行过滤,对每个元素应用f函数,返回值为true的元素在RDD中保留,返回为false的将过滤掉。 内部实现相当于生成FilteredRDD(this,sc.clean(f))。

图中,每个方框代表一个RDD分区。 T可以是任意的类型。通过用户自定义的过滤函数f,对每个数据项进行操作,将满足条件,返回结果为true的数据项保留。 例如,过滤掉V2、 V3保留了V1,将区分命名为V1’。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return a new RDD containing only the elements that satisfy a predicate.  
  3.  */  
  4. def filter(f: T => Boolean): RDD[T] = {  
  5.   val cleanF = sc.clean(f)  
  6.   new MapPartitionsRDD[T, T](  
  7.     this,  
  8.     (context, pid, iter) => iter.filter(cleanF),  
  9.     preservesPartitioning = true)  
  10. }  

(2)distinct

distinct将RDD中的元素进行去重操作。 

图中,每个方框代表一个分区,通过distinct函数,将数据去重。 例如,重复数据V1、 V1去重后只保留一份V1。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return a new RDD containing the distinct elements in this RDD.  
  3.  */  
  4. def distinct(numPartitions: Int)(implicit ord: Ordering[T] = null): RDD[T] =  
  5.   map(x => (x, null)).reduceByKey((x, y) => x, numPartitions).map(_._1)  
  6.   
  7. /**  
  8.  * Return a new RDD containing the distinct elements in this RDD.  
  9.  */  
  10. def distinct(): RDD[T] = distinct(partitions.size)  

(3)subtract

subtract相当于进行集合的差操作,RDD 1去除RDD 1和RDD 2交集中的所有元素。 

图中,左侧的大方框代表两个RDD,大方框内的小方框代表RDD的分区。右侧大方框代表合并后的RDD,大方框内的小方框代表分区。V1在两个RDD中均有,根据差集运算规则,新RDD不保留,V2在第一个RDD有,第二个RDD没有,则在新RDD元素中包含V2。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return an RDD with the elements from `this` that are not in `other`.  
  3.  *  
  4.  * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting  
  5.  * RDD will be <= us.  
  6.  */  
  7. def subtract(other: RDD[T]): RDD[T] =  
  8.   subtract(other, partitioner.getOrElse(new HashPartitioner(partitions.size)))  
  9.   
  10. /**  
  11.  * Return an RDD with the elements from `this` that are not in `other`.  
  12.  */  
  13. def subtract(other: RDD[T], numPartitions: Int): RDD[T] =  
  14.   subtract(other, new HashPartitioner(numPartitions))  
  15.   
  16. /**  
  17.  * Return an RDD with the elements from `this` that are not in `other`.  
  18.  */  
  19. def subtract(other: RDD[T], p: Partitioner)(implicit ord: Ordering[T] = null): RDD[T] = {  
  20.   if (partitioner == Some(p)) {  
  21.     // Our partitioner knows how to handle T (which, since we have a partitioner, is  
  22.     // really (K, V)) so make a new Partitioner that will de-tuple our fake tuples  
  23.     val p2 = new Partitioner() {  
  24.       override def numPartitions = p.numPartitions  
  25.       override def getPartition(k: Any) = p.getPartition(k.asInstanceOf[(Any, _)]._1)  
  26.     }  
  27.     // Unfortunately, since we're making a new p2, we'll get ShuffleDependencies  
  28.     // anyway, and when calling .keys, will not have a partitioner set, even though  
  29.     // the SubtractedRDD will, thanks to p2's de-tupled partitioning, already be  
  30.     // partitioned by the right/real keys (e.g. p).  
  31.     this.map(x => (x, null)).subtractByKey(other.map((_, null)), p2).keys  
  32.   } else {  
  33.     this.map(x => (x, null)).subtractByKey(other.map((_, null)), p).keys  
  34.   }  
  35. }  

(4)sample

sample将RDD这个集合内的元素进行采样,获取所有元素的子集。用户可以设定是否有放回的抽样、百分比、随机种子,进而决定采样方式。 
参数说明:

withReplacement=true, 表示有放回的抽样; 
withReplacement=false, 表示无放回的抽样。


每个方框是一个RDD分区。通过sample函数,采样50%的数据。V1、V2、U1、U2、U3、U4采样出数据V1和U1、U2,形成新的RDD。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return a sampled subset of this RDD.  
  3.  */  
  4. def sample(withReplacement: Boolean,  
  5.     fraction: Double,  
  6.     seed: Long = Utils.random.nextLong): RDD[T] = {  
  7.   require(fraction >= 0.0, "Negative fraction value: " + fraction)  
  8.   if (withReplacement) {  
  9.     new PartitionwiseSampledRDD[T, T](this, new PoissonSampler[T](fraction), true, seed)  
  10.   } else {  
  11.     new PartitionwiseSampledRDD[T, T](this, new BernoulliSampler[T](fraction), true, seed)  
  12.   }  
  13. }  

(5)takeSample

takeSample()函数和上面的sample函数是一个原理,但是不使用相对比例采样,而是按设定的采样个数进行采样,同时返回结果不再是RDD,而是相当于对采样后的数据进行collect(),返回结果的集合为单机的数组。

图中,左侧的方框代表分布式的各个节点上的分区,右侧方框代表单机上返回的结果数组。通过takeSample对数据采样,设置为采样一份数据,返回结果为V1。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Return a fixed-size sampled subset of this RDD in an array  
  3.  *  
  4.  * @param withReplacement whether sampling is done with replacement  
  5.  * @param num size of the returned sample  
  6.  * @param seed seed for the random number generator  
  7.  * @return sample of specified size in an array  
  8.  */  
  9. def takeSample(withReplacement: Boolean,  
  10.     num: Int,  
  11.     seed: Long = Utils.random.nextLong): Array[T] = {  
  12.   val numStDev =  10.0  
  13.   
  14.   if (num < 0) {  
  15.     throw new IllegalArgumentException("Negative number of elements requested")  
  16.   } else if (num == 0) {  
  17.     return new Array[T](0)  
  18.   }  
  19.   
  20.   val initialCount = this.count()  
  21.   if (initialCount == 0) {  
  22.     return new Array[T](0)  
  23.   }  
  24.   
  25.   val maxSampleSize = Int.MaxValue - (numStDev * math.sqrt(Int.MaxValue)).toInt  
  26.   if (num > maxSampleSize) {  
  27.     throw new IllegalArgumentException("Cannot support a sample size > Int.MaxValue - " +  
  28.       s"$numStDev * math.sqrt(Int.MaxValue)")  
  29.   }  
  30.   
  31.   val rand = new Random(seed)  
  32.   if (!withReplacement && num >= initialCount) {  
  33.     return Utils.randomizeInPlace(this.collect(), rand)  
  34.   }  
  35.   
  36.   val fraction = SamplingUtils.computeFractionForSampleSize(num, initialCount,  
  37.     withReplacement)  
  38.   
  39.   var samples = this.sample(withReplacement, fraction, rand.nextInt()).collect()  
  40.   
  41.   // If the first sample didn't turn out large enough, keep trying to take samples;  
  42.   // this shouldn't happen often because we use a big multiplier for the initial size  
  43.   var numIters = 0  
  44.   while (samples.length < num) {  
  45.     logWarning(s"Needed to re-sample due to insufficient sample size. Repeat #$numIters")  
  46.     samples = this.sample(withReplacement, fraction, rand.nextInt()).collect()  
  47.     numIters += 1  
  48.   }  
  49.   
  50.   Utils.randomizeInPlace(samples, rand).take(num)  
  51. }  

Cache型

(1)cache

cache将RDD元素从磁盘缓存到内存,相当于persist(MEMORY_ONLY)函数的功能。 

图中,每个方框代表一个RDD分区,左侧相当于数据分区都存储在磁盘,通过cache算子将数据缓存在内存。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */  
  2. def cache(): this.type = persist()  

(2)persist

persist函数对RDD进行缓存操作。数据缓存在哪里由StorageLevel枚举类型确定。 
有几种类型的组合,DISK代表磁盘,MEMORY代表内存,SER代表数据是否进行序列化存储。StorageLevel是枚举类型,代表存储模式,如,MEMORY_AND_DISK_SER代表数据可以存储在内存和磁盘,并且以序列化的方式存储。 其他同理。



图中,方框代表RDD分区。 disk代表存储在磁盘,mem代表存储在内存。 数据最初全部存储在磁盘,通过persist(MEMORY_AND_DISK)将数据缓存到内存,但是有的分区无法容纳在内存,例如:图3-18中将含有V1,V2,V3的RDD存储到磁盘,将含有U1,U2的RDD仍旧存储在内存。

源码:

[plain] view plain copy
 
 在CODE上查看代码片派生到我的代码片
  1. /**  
  2.  * Set this RDD's storage level to persist its values across operations after the first time  
  3.  * it is computed. This can only be used to assign a new storage level if the RDD does not  
  4.  * have a storage level set yet..  
  5.  */  
  6. def persist(newLevel: StorageLevel): this.type = {  
  7.   // TODO: Handle changes of StorageLevel  
  8.   if (storageLevel != StorageLevel.NONE && newLevel != storageLevel) {  
  9.     throw new UnsupportedOperationException(  
  10.       "Cannot change storage level of an RDD after it was already assigned a level")  
  11.   }  
  12.   sc.persistRDD(this)  
  13.   // Register the RDD with the ContextCleaner for automatic GC-based cleanup  
  14.   sc.cleaner.foreach(_.registerRDDForCleanup(this))  
  15.   storageLevel = newLevel  
  16.   this  
  17. }  
  18.   
  19. /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */  
  20. def persist(): this.type = persist(StorageLevel.MEMORY_ONLY)  

原文链接:http://blog.csdn.net/jasonding1354

原文地址:https://www.cnblogs.com/duanxz/p/6327370.html