和一般RDD最大的不同就是有两个泛型参数, [K, V]表示pair的概念
关键的function是, combineByKey, 所有pair相关操作的抽象
combine是这样的操作, Turns an RDD[(K, V)] into a result of type RDD[(K, C)]
其中C有可能只是简单类型, 但经常是seq, 比如(Int, Int) to (Int, Seq[Int])
下面来看看combineByKey的参数,
首先需要用户自定义一些操作,
createCombiner: V => C, C不存在的情况下, 比如通过V创建seq C
mergeValue: (C, V) => C, 当C已经存在的情况下, 需要merge, 比如把item V加到seq C中, 或者叠加
mergeCombiners: (C, C) => C, 合并两个C
partitioner: Partitioner, Shuffle时需要的Partitioner
mapSideCombine: Boolean = true, 为了减小传输量, 很多combine可以在map端先做, 比如叠加, 可以先在一个partition中把所有相同的key的value叠加, 再shuffle
serializerClass: String = null, 传输需要序列化, 用户可以自定义序列化类
/** * Extra functions available on RDDs of (key, value) pairs through an implicit conversion. * Import `org.apache.spark.SparkContext._` at the top of your program to use these functions. */ class PairRDDFunctions[K: ClassManifest, V: ClassManifest](self: RDD[(K, V)]) extends Logging with SparkHadoopMapReduceUtil with Serializable { /** * Generic function to combine the elements for each key using a custom set of aggregation * functions. Turns an RDD[(K, V)] into a result of type RDD[(K, C)], for a "combined type" C * Note that V and C can be different -- for example, one might group an RDD of type * (Int, Int) into an RDD of type (Int, Seq[Int]). Users provide three functions: * * - `createCombiner`, which turns a V into a C (e.g., creates a one-element list) * - `mergeValue`, to merge a V into a C (e.g., adds it to the end of a list) * - `mergeCombiners`, to combine two C's into a single one. * * In addition, users can control the partitioning of the output RDD, and whether to perform * map-side aggregation (if a mapper can produce multiple items with the same key). */ def combineByKey[C](createCombiner: V => C, mergeValue: (C, V) => C, mergeCombiners: (C, C) => C, partitioner: Partitioner, mapSideCombine: Boolean = true, serializerClass: String = null): RDD[(K, C)] = {
val aggregator = new Aggregator[K, V, C](createCombiner, mergeValue, mergeCombiners) //1.Aggregator
//RDD本身的partitioner和传入的partitioner相等时, 即不需要重新shuffle, 直接map即可 if (self.partitioner == Some(partitioner)) { self.mapPartitions(aggregator.combineValuesByKey, preservesPartitioning = true) //2. mapPartitions, map端直接调用combineValuesByKey } else if (mapSideCombine) { //如果需要mapSideCombine val combined = self.mapPartitions(aggregator.combineValuesByKey, preservesPartitioning = true) //先在partition内部做mapSideCombine val partitioned = new ShuffledRDD[K, C, (K, C)](combined, partitioner).setSerializer(serializerClass) //3. ShuffledRDD, 进行shuffle partitioned.mapPartitions(aggregator.combineCombinersByKey, preservesPartitioning = true) //Shuffle完后, 在reduce端再做一次combine, 使用combineCombinersByKey } else { // Don't apply map-side combiner.和上面的区别就是不做mapSideCombine // A sanity check to make sure mergeCombiners is not defined. assert(mergeCombiners == null) val values = new ShuffledRDD[K, V, (K, V)](self, partitioner).setSerializer(serializerClass) values.mapPartitions(aggregator.combineValuesByKey, preservesPartitioning = true) } } }
1. Aggregator
在combineByKey中, 首先创建Aggregator, 其实在Aggregator中封装了两个函数,
combineValuesByKey, 用于处理将V加入到C的case, 比如加入一个item到一个seq里面, 用于map端
combineCombinersByKey, 用于处理两个C合并, 比如两个seq合并, 用于reduce端
case class Aggregator[K, V, C] ( createCombiner: V => C, mergeValue: (C, V) => C, mergeCombiners: (C, C) => C) { def combineValuesByKey(iter: Iterator[_ <: Product2[K, V]]) : Iterator[(K, C)] = { val combiners = new JHashMap[K, C] for (kv <- iter) { val oldC = combiners.get(kv._1) if (oldC == null) { combiners.put(kv._1, createCombiner(kv._2)) } else { combiners.put(kv._1, mergeValue(oldC, kv._2)) } } combiners.iterator } def combineCombinersByKey(iter: Iterator[(K, C)]) : Iterator[(K, C)] = { val combiners = new JHashMap[K, C] iter.foreach { case(k, c) => val oldC = combiners.get(k) if (oldC == null) { combiners.put(k, c) } else { combiners.put(k, mergeCombiners(oldC, c)) } } combiners.iterator } }
2. mapPartitions
mapPartitions其实就是使用MapPartitionsRDD
做的事情就是对当前partition执行map函数f, Iterator[T] => Iterator[U]
比如, 执行combineValuesByKey: Iterator[_ <: Product2[K, V]] to Iterator[(K, C)]
/** * Return a new RDD by applying a function to each partition of this RDD. */ def mapPartitions[U: ClassManifest](f: Iterator[T] => Iterator[U], preservesPartitioning: Boolean = false): RDD[U] = new MapPartitionsRDD(this, sc.clean(f), preservesPartitioning)
private[spark] class MapPartitionsRDD[U: ClassManifest, T: ClassManifest]( prev: RDD[T], f: Iterator[T] => Iterator[U], preservesPartitioning: Boolean = false) extends RDD[U](prev) { override val partitioner = if (preservesPartitioning) firstParent[T].partitioner else None override def getPartitions: Array[Partition] = firstParent[T].partitions override def compute(split: Partition, context: TaskContext) = f(firstParent[T].iterator(split, context)) // 对于map,就是调用f
3. ShuffledRDD
Shuffle实际上是由系统的shuffleFetcher完成的, Spark的抽象封装非常的好
所以在这里看不到Shuffle具体是怎么样做的, 这个需要分析到shuffleFetcher时候才能看到
因为每个shuffle是有一个全局的shuffleid的
所以在compute里面, 你只是看到用BlockStoreShuffleFetcher根据shuffleid和partitionid直接fetch到shuffle过后的数据
/** * The resulting RDD from a shuffle (e.g. repartitioning of data). * @param prev the parent RDD. * @param part the partitioner used to partition the RDD * @tparam K the key class. * @tparam V the value class. */ class ShuffledRDD[K, V, P <: Product2[K, V] : ClassManifest]( @transient var prev: RDD[P], part: Partitioner) extends RDD[P](prev.context, Nil) {
override val partitioner = Some(part)
//ShuffleRDD会进行repartition, 所以从Partitioner中取出新的part数目
//并用Array.tabulate动态创建相应个数的ShuffledRDDPartition override def getPartitions: Array[Partition] = { Array.tabulate[Partition](part.numPartitions)(i => new ShuffledRDDPartition(i)) } override def compute(split: Partition, context: TaskContext): Iterator[P] = { val shuffledId = dependencies.head.asInstanceOf[ShuffleDependency[K, V]].shuffleId SparkEnv.get.shuffleFetcher.fetch[P](shuffledId, split.index, context.taskMetrics, SparkEnv.get.serializerManager.get(serializerClass)) } }
ShuffledRDDPartition没啥区别, 一样只是记录id
private[spark] class ShuffledRDDPartition(val idx: Int) extends Partition { override val index = idx override def hashCode(): Int = idx }
下面再来看一下, 如果使用combineByKey来实现其他的操作的,
group
group是比较典型的例子, (Int, Int) to (Int, Seq[Int])
由于groupByKey不使用map side combine, 因为这样也无法减少传输空间, 所以不需要实现mergeCombiners
/** * Group the values for each key in the RDD into a single sequence. Allows controlling the * partitioning of the resulting key-value pair RDD by passing a Partitioner. */ def groupByKey(partitioner: Partitioner): RDD[(K, Seq[V])] = { // groupByKey shouldn't use map side combine because map side combine does not // reduce the amount of data shuffled and requires all map side data be inserted // into a hash table, leading to more objects in the old gen. def createCombiner(v: V) = ArrayBuffer(v) //创建seq def mergeValue(buf: ArrayBuffer[V], v: V) = buf += v //添加item到seq val bufs = combineByKey[ArrayBuffer[V]]( createCombiner _, mergeValue _, null, partitioner, mapSideCombine=false) bufs.asInstanceOf[RDD[(K, Seq[V])]] }
reduce
reduce是更简单的一种情况, 只是两个值合并成一个值, (Int, Int V) to (Int, Int C), 比如叠加
所以createCombiner很简单, 就是直接返回v
而mergeValue和mergeCombiners逻辑是相同的, 没有区别
/**
* Merge the values for each key using an associative reduce function. This will also perform
* the merging locally on each mapper before sending results to a reducer, similarly to a
* "combiner" in MapReduce.
*/
def reduceByKey(partitioner: Partitioner, func: (V, V) => V): RDD[(K, V)] = {
combineByKey[V]((v: V) => v, func, func, partitioner)
}
本文章摘自博客园，原文发布日期：2013-12-24
