Spark RDD概念学习系列之rdd持久化、广播、累加器（十八）
1、rdd持久化 2、广播 3、累加器 1、rdd持久化 通过spark-shell，可以快速的验证我们的想法和操作！ 启动hdfs集群 spark@SparkSingleNode:/usr/local/hadoop/hadoop-2.6.0$sbin/start-dfs.sh 启动spark集群 spark@SparkSingleNode:/usr/local/spark/spark-1.5.2-bin-hadoop2.6$sbin/start-all.sh 启动spark-shell spark@SparkSingleNode:/usr/local/spark/spark-1.5.2-bin-hadoop2.6/bin$./spark-shell --master spark://SparkSingleNode:7077 --executor-memory 1g reduce scala>sc res0: org.apache.spark.SparkContext = org.apache.spark.SparkContext@3bcc8f13 scala> val numbers = sc.parallelize <console>:21: error: missing arguments for method parallelize in class SparkContext; follow this method with `_' if you want to treat it as a partially applied function val numbers = sc.parallelize ^ scala>val numbers = sc.parallelize(1 to 100) numbers: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[0] at parallelize at <console>:21 scala>numbers.reduce(_+_) took 11.790246 s res1: Int = 5050 可见，reduce是个action。 scala>val result = numbers.map(2*_) result: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[1] at map at <console>:23 scala>val data = result.collect reduce源码 /** * Reduces the elements of this RDD using the specified commutative and * associative binary operator. */ def reduce(f: (T, T) => T): T = withScope { val cleanF = sc.clean(f) val reducePartition: Iterator[T] => Option[T] = iter => { if (iter.hasNext) { Some(iter.reduceLeft(cleanF)) } else { None } } var jobResult: Option[T] = None val mergeResult = (index: Int, taskResult: Option[T]) => { if (taskResult.isDefined) { jobResult = jobResult match { case Some(value) => Some(f(value, taskResult.get)) case None => taskResult } } } sc.runJob(this, reducePartition, mergeResult) // Get the final result out of our Option, or throw an exception if the RDD was empty jobResult.getOrElse(throw new UnsupportedOperationException("empty collection")) } 可见,这也是一个action操作。 collect data: Array[Int] = Array(2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 62, 64, 66, 68, 70, 72, 74, 76, 78, 80, 82, 84, 86, 88, 90, 92, 94, 96, 98, 100, 102, 104, 106, 108, 110, 112, 114, 116, 118, 120, 122, 124, 126, 128, 130, 132, 134, 136, 138, 140, 142, 144, 146, 148, 150, 152, 154, 156, 158, 160, 162, 164, 166, 168, 170, 172, 174, 176, 178, 180, 182, 184, 186, 188, 190, 192, 194, 196, 198, 200) scala> collect源码 /** * Return an array that contains all of the elements in this RDD. */ def collect(): Array[T] = withScope { val results = sc.runJob(this, (iter: Iterator[T]) => iter.toArray) Array.concat(results: _*) } 可见,这也是一个action操作。 从收集结果的角度来说，如果想要在命令行终端中，看到执行结果，就必须collect。 从源码的角度来说，凡是action级别的操作，都会触发sc.rubJob。这点，spark里是一个应用程序允许有多个Job,而hadoop里一个应用程序只能一个Job。 count scala>numbers res2: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[0] at parallelize at <console>:21 scala>1 to 100 res3: scala.collection.immutable.Range.Inclusive = Range(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100) scala>numbers.count took 0.649005 s res4: Long = 100 count源码 /** * Return the number of elements in the RDD. */ def count(): Long = sc.runJob(this, Utils.getIteratorSize _).sum 可见,这也是一个action操作。 take scala>val topN = numbers.take(5) topN: Array[Int] = Array(1, 2, 3, 4, 5) take源码 /** * Take the first num elements of the RDD. It works by first scanning one partition, and use the * results from that partition to estimate the number of additional partitions needed to satisfy * the limit. * * @note due to complications in the internal implementation, this method will raise * an exception if called on an RDD of `Nothing` or `Null`. */ def take(num: Int): Array[T] = withScope { if (num == 0) { new Array[T](0) } else { val buf = new ArrayBuffer[T] val totalParts = this.partitions.length var partsScanned = 0 while (buf.size < num && partsScanned < totalParts) { // The number of partitions to try in this iteration. It is ok for this number to be // greater than totalParts because we actually cap it at totalParts in runJob. var numPartsToTry = 1 if (partsScanned > 0) { // If we didn't find any rows after the previous iteration, quadruple and retry. // Otherwise, interpolate the number of partitions we need to try, but overestimate // it by 50%. We also cap the estimation in the end. if (buf.size == 0) { numPartsToTry = partsScanned * 4 } else { // the left side of max is >=1 whenever partsScanned >= 2 numPartsToTry = Math.max((1.5 * num * partsScanned / buf.size).toInt - partsScanned, 1) numPartsToTry = Math.min(numPartsToTry, partsScanned * 4) } } val left = num - buf.size val p = partsScanned until math.min(partsScanned + numPartsToTry, totalParts) val res = sc.runJob(this, (it: Iterator[T]) => it.take(left).toArray, p) res.foreach(buf ++= _.take(num - buf.size)) partsScanned += numPartsToTry } buf.toArray } } 可见,这也是一个action操作。 countByKey scala>val scores = Array(Tuple2(1,100),Tuple2(1,100),Tuple2(2,100),Tuple2(2,100),Tuple2(3,100)) scores: Array[(Int, Int)] = Array((1,100), (1,100), (2,100), (2,100), (3,100)) scala> val content = sc.parallelize <console>:21: error: missing arguments for method parallelize in class SparkContext; follow this method with `_' if you want to treat it as a partially applied function val content = sc.parallelize ^ scala>val content = sc.parallelize(scores) content: org.apache.spark.rdd.RDD[(Int, Int)] = ParallelCollectionRDD[0] at parallelize at <console>:23 scala>val data = content.countByKey() took 10.556634 s data: scala.collection.Map[Int,Long] = Map(2 -> 2, 1 -> 2, 3 -> 1) countByKey源码 /** * Count the number of elements for each key, collecting the results to a local Map. * * Note that this method should only be used if the resulting map is expected to be small, as * the whole thing is loaded into the driver's memory. * To handle very large results, consider using rdd.mapValues(_ => 1L).reduceByKey(_ + _), which * returns an RDD[T, Long] instead of a map. */ def countByKey(): Map[K, Long] = self.withScope { self.mapValues(_ => 1L).reduceByKey(_ + _).collect().toMap } 可见,这也是一个action操作。 saveAsTextFile 之前，在 rdd实战（rdd基本操作实战及transformation和action流程图）（源码） scala>val partitionsReadmeRdd = sc.textFile("hdfs://SparkSingleNode:9000/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).saveAsTextFile("~/partition1README.txt") 这里呢。 scala>val partitionsReadmeRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).saveAsTextFile("/partition1README.txt") scala>val partitionsReadmeRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).saveAsTextFile("/partition1README.txt") saveAsTextFile源码 /** * Save this RDD as a text file, using string representations of elements. */ def saveAsTextFile(path: String): Unit = withScope { // https://issues.apache.org/jira/browse/SPARK-2075 // // NullWritable is a `Comparable` in Hadoop 1.+, so the compiler cannot find an implicit // Ordering for it and will use the default `null`. However, it's a `Comparable[NullWritable]` // in Hadoop 2.+, so the compiler will call the implicit `Ordering.ordered` method to create an // Ordering for `NullWritable`. That's why the compiler will generate different anonymous // classes for `saveAsTextFile` in Hadoop 1.+ and Hadoop 2.+. // // Therefore, here we provide an explicit Ordering `null` to make sure the compiler generate // same bytecodes for `saveAsTextFile`. val nullWritableClassTag = implicitly[ClassTag[NullWritable]] val textClassTag = implicitly[ClassTag[Text]] val r = this.mapPartitions { iter => val text = new Text() iter.map { x => text.set(x.toString) (NullWritable.get(), text) } } RDD.rddToPairRDDFunctions(r)(nullWritableClassTag, textClassTag, null) .saveAsHadoopFile[TextOutputFormat[NullWritable, Text]](path) } /** * Save this RDD as a compressed text file, using string representations of elements. */ def saveAsTextFile(path: String, codec: Class[_ <: CompressionCodec]): Unit = withScope { // https://issues.apache.org/jira/browse/SPARK-2075 val nullWritableClassTag = implicitly[ClassTag[NullWritable]] val textClassTag = implicitly[ClassTag[Text]] val r = this.mapPartitions { iter => val text = new Text() iter.map { x => text.set(x.toString) (NullWritable.get(), text) } } RDD.rddToPairRDDFunctions(r)(nullWritableClassTag, textClassTag, null) .saveAsHadoopFile[TextOutputFormat[NullWritable, Text]](path, codec) } saveAsTextFile不仅，可保存在集群里，也可以保存到本地，这就要看hadoop的运行模式。 由此可见，它也是个action操作。 以上是rdd持久化的第一个方面，就是action级别的操作。 rdd持久化的第二个方面，就是通过persist。为什么在spark里，随处可见persist的身影呢？ 原因一：spark在默认情况下，数据是放在内存中，适合高速迭代。比如在一个stage里，有1000个步骤，它其实只在第1个步骤输入数据，在第1000个步骤输出数据，在中间不产生临时数据。但是，分布式系统，分享非常高，所以，容出错，设计到容错。 由于，rdd是有血统继承关系的，即lineager。如果后面的rdd数据分片出错了或rdd本身出错了，则，可根据其前面依赖的lineager，算出来。 但是，假设1000个步骤，如果之前，没有父rdd进行persist或cache的话，则要重头开始了。亲！ 什么时候，该persist？ 1、在某个步骤非常费时的情况下，不好使 （手动） 2、计算链条特别长的情况下 （手动） 3、checkpoint所在的rdd也一定要持久化数据 （注意：在checkpoint之前，进行persist） （手动） checkpoint是rdd的算子， 先写，某个具体rdd.checkpoint 或 某个具体rdd.cache ，再写， 某个具体rdd.persist 4、shuffle之后 （因为shuffle之后，要网络传输，风险大） （手动） 5、shuffle之前 （框架，默认给我们做的，把数据持久化到本地磁盘） checkpoint源码 /** * Mark this RDD for checkpointing. It will be saved to a file inside the checkpoint * directory set with `SparkContext#setCheckpointDir` and all references to its parent * RDDs will be removed. This function must be called before any job has been * executed on this RDD. It is strongly recommended that this RDD is persisted in * memory, otherwise saving it on a file will require recomputation. */ def checkpoint(): Unit = RDDCheckpointData.synchronized { // NOTE: we use a global lock here due to complexities downstream with ensuring // children RDD partitions point to the correct parent partitions. In the future // we should revisit this consideration. if (context.checkpointDir.isEmpty) { throw new SparkException("Checkpoint directory has not been set in the SparkContext") } else if (checkpointData.isEmpty) { checkpointData = Some(new ReliableRDDCheckpointData(this)) } } persist源码 /** * Mark this RDD for persisting using the specified level. * * @param newLevel the target storage level * @param allowOverride whether to override any existing level with the new one */ private def persist(newLevel: StorageLevel, allowOverride: Boolean): this.type = { // TODO: Handle changes of StorageLevel if (storageLevel != StorageLevel.NONE && newLevel != storageLevel && !allowOverride) { throw new UnsupportedOperationException( "Cannot change storage level of an RDD after it was already assigned a level") } // If this is the first time this RDD is marked for persisting, register it // with the SparkContext for cleanups and accounting. Do this only once. if (storageLevel == StorageLevel.NONE) { sc.cleaner.foreach(_.registerRDDForCleanup(this)) sc.persistRDD(this) } storageLevel = newLevel this } /** * Set this RDD's storage level to persist its values across operations after the first time * it is computed. This can only be used to assign a new storage level if the RDD does not * have a storage level set yet. Local checkpointing is an exception. */ def persist(newLevel: StorageLevel): this.type = { if (isLocallyCheckpointed) { // This means the user previously called localCheckpoint(), which should have already // marked this RDD for persisting. Here we should override the old storage level with // one that is explicitly requested by the user (after adapting it to use disk). persist(LocalRDDCheckpointData.transformStorageLevel(newLevel), allowOverride = true) } else { persist(newLevel, allowOverride = false) } } /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */ def persist(): this.type = persist(StorageLevel.MEMORY_ONLY) /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */ def cache(): this.type = persist() StorageLevel里有很多类型 这里，牵扯到序列化。 问，为什么要序列化？ 答：节省空间，减少体积。内存不够时，把MEMORY中的数据，进行序列化。 当然，也有不好一面，序列化时，会反序列化，反序列化耗cpu。 MEMORY_AND_DISK 假设，我们制定数据存储方式是，MEMORY_AND_DISK。则，是不是同时，存储到内存和磁盘呢？ 答：不是啊，亲。spark一定是优先考虑内存的啊，只要内存足够，不会考虑磁盘。若内存不够了，则才放部分数据到磁盘。 极大地减少数据丢失概率发生。 MEMORY_ONLY 假设，我们制定数据存储方式是，MEMORY_ONLY。则，只放到内存。当内存不够了，会出现OOM。或数据丢失。 OFF_HEAP 这牵扯到Tachyon，基于内存的分布式系统 为什么有2分副本？好处是？ 假设，一个计算特别耗时，而且，又是基于内存，如果其中一份副本崩溃掉，则可迅速切换到另一份副本去计算。这就是“空间换时间”！非常重要 这不是并行计算，这是计算后的结果，放2份副本。 scala>val partitionsReadmeRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).count took 6.270138 s scala>val partitionsReadmeRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache.count took 4.147545 s scala>val partitionsReadmeRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache.count took 4.914212 s scala>val cacheRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache scala>cacheRdd.count took 3.371621 scala>val cacheRdd = sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache scala>cacheRdd.count took 0.943499 s 我的天啊！ scala>sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache.count took 5.603903 scala>sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache.count took 4.146627 scala>sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).cache.count took 3.071122 cache之后，一定不能立即有其他算子！ 实际工程中， cache之后，如果有其他算子，则会，重新触发这个工作过程。 注意:cache，不是action cache缓存，怎么让它失效？ 答：unpersist persist是lazy级别的，unpersist是eager级别的。cache是persist的一个特殊情况。 cache和persist的区别？ 答：persist可以放到磁盘、放到内存、同时放到内存和磁盘。以及多份副本 cache只能放到内存，以及只能一份副本。 persisit在内存不够时，保存在磁盘的哪个目录？ 答：local的process。 好的，以上是，rdd持久化的两个方面。 rdd持久化的第一个方面，就是常用的action级别的操作。 rdd持久化的第二个方面，就是持久化的不同方式，以及它内部的运行情况 小知识：cache之后，一定不能立即有其他算子！实际工程中， cache之后，如果有其他算子，则会，重新触发这个工作过程。 一般都不会跨机器抓内存，宁愿排队。宁愿数据不动代码动。 2、广播 为什么要有，rdd广播？ 答：大变量、join、冗余、减少数据移动、通信、状态、集群消息、共享、网络传输慢要提前、数据量大耗网络、减少通信、要同步。 为什么大变量，需要广播呢？ 答：原因是，每个task运行，读取全集数据时，task本身执行时，每次都要拷贝一份数据副本，如果变量比较大，如一百万，则要拷贝一百万。 （2）广播（线程中共享）不必每个task都拷贝一份副本，因为它是全局唯一的，极大的减少oom，减少通信，冗余、共享变量等。广播是将数据广播到Executor的内存中，其内部所以的任务都会只享有全局唯一的变量，减少网络传输。 text在读取数据时候，拷贝一份的数据副本（变量），因为函数式编程（变量不变），不拷贝状态容易被改变，数据量小（1、引用较小2、数据本身小），变量大容易产生oom（task拷贝数据 在内存中运行），网络传输慢，要提前，冗余、共享，减少通信。 广播变量： 广播变量允许程序员将一个只读的变量缓存在每台机器上，而不用在任务之间传递变量。广播变量可被用于有效地给每个节点一个大输入数据集的副本。Spark还尝试使用高效地广播算法来分发变量，进而减少通信的开销。 Spark的动作通过一系列的步骤执行，这些步骤由分布式的洗牌操作分开。Spark自动地广播每个步骤每个任务需要的通用数据。这些广播数据被序列化地缓存，在运行任务之前被反序列化出来。这意味着当我们需要在多个阶段的任务之间使用相同的数据，或者以反序列化形式缓存数据是十分重要的时候，显式地创建广播变量才有用。 （本段摘自：http://blog.csdn.net/happyanger6/article/details/46576831） 广播工作机制图 广播工作机制图 参考： http://blog.csdn.net/kxr0502/article/details/50574561 问：读广播，会消耗网络传输吗？ 答：不消耗，广播是放在内存中。读取它，不消耗。 问：广播变量是不是就是向每一个executor，广播一份数据，而不是向每一个task，广播一份数据？这样对吗? 答：对 广播是由Driver发给当前Application分配的所有Executor内存级别的全局只读变量，Executor中的线程池中的线程共享该全局变量，极大的减少了网络传输（否则的话每个Task都要传输一次该变量）并极大的节省了内存，当然也隐形的提高的CPU的有效工作。 实战创建广播： scala>val number = 10 number: Int = 10 scala>val broadcastNumber = sc.broadcast(number) 16/09/29 17:26:47 INFO storage.MemoryStore: ensureFreeSpace(40) called with curMem=1782734, maxMem=560497950 16/09/29 17:26:47 INFO storage.MemoryStore: Block broadcast_38 stored as values in memory (estimated size 40.0 B, free 532.8 MB) 16/09/29 17:26:48 INFO storage.MemoryStore: ensureFreeSpace(97) called with curMem=1782774, maxMem=560497950 16/09/29 17:26:48 INFO storage.MemoryStore: Block broadcast_38_piece0 stored as bytes in memory (estimated size 97.0 B, free 532.8 MB) 16/09/29 17:26:48 INFO storage.BlockManagerInfo: Added broadcast_38_piece0 in memory on 192.168.80.128:40914 (size: 97.0 B, free: 534.4 MB) 16/09/29 17:26:48 INFO spark.SparkContext: Created broadcast 38 from broadcast at <console>:23broadcastNumber: org.apache.spark.broadcast.Broadcast[Int] = Broadcast(38) scala> val data = sc.parallelize <console>:21: error: missing arguments for method parallelize in class SparkContext; follow this method with `_' if you want to treat it as a partially applied function val data = sc.parallelize ^ scala>val data = sc.parallelize(1 to 100) data: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[61] at parallelize at <console>:21 scala>val bn = data.map(_* broadcastNumber.value) bn: org.apache.spark.rdd.RDD[Int] = MapPartitionsRDD[62] at map at <console>:27 scala> 我们知道，是test是要广播变量，但，我们编程，对rdd。 //通过在一个变量v上调用SparkContext.broadcast(v)可以创建广播变量。广播变量是围绕着v的封装，可以通过value方法访问这个变量。 问：广播变量里有很多变量吗？ 答:当然可以有很多，用java bin或scala封装，就可以了。 如，在这里。广播变量是，broadcastNumber， 里，有变量value等。 scala>val broadcastNumber = sc.broadcast(number) scala>val bn = data.map(_* broadcastNumber.value) scala>bn.collect res12: Array[Int] = Array(10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, 400, 410, 420, 430, 440, 450, 460, 470, 480, 490, 500, 510, 520, 530, 540, 550, 560, 570, 580, 590, 600, 610, 620, 630, 640, 650, 660, 670, 680, 690, 700, 710, 720, 730, 740, 750, 760, 770, 780, 790, 800, 810, 820, 830, 840, 850, 860, 870, 880, 890, 900, 910, 920, 930, 940, 950, 960, 970, 980, 990, 1000) scala> 由此，可见，通过机制、流程图和实战，深度剖析对广播全面详解！ broadcast源码分析 参考：http://www.cnblogs.com/seaspring/p/5682053.html BroadcastManager源码 /* * Licensed to the Apache Software Foundation (ASF) under one or more * contributor license agreements. See the NOTICE file distributed with * this work for additional information regarding copyright ownership. * The ASF licenses this file to You under the Apache License, Version 2.0 * (the "License"); you may not use this file except in compliance with * the License. You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ package org.apache.spark.broadcast import java.util.concurrent.atomic.AtomicLong import scala.reflect.ClassTag import org.apache.spark._ import org.apache.spark.util.Utils private[spark] class BroadcastManager( val isDriver: Boolean, conf: SparkConf, securityManager: SecurityManager) extends Logging { private var initialized = false private var broadcastFactory: BroadcastFactory = null initialize() // Called by SparkContext or Executor before using Broadcast private def initialize() { synchronized { if (!initialized) { val broadcastFactoryClass = conf.get("spark.broadcast.factory", "org.apache.spark.broadcast.TorrentBroadcastFactory") broadcastFactory = Utils.classForName(broadcastFactoryClass).newInstance.asInstanceOf[BroadcastFactory] // Initialize appropriate BroadcastFactory and BroadcastObject broadcastFactory.initialize(isDriver, conf, securityManager) initialized = true } } } def stop() { broadcastFactory.stop() } private val nextBroadcastId = new AtomicLong(0) def newBroadcast[T: ClassTag](value_ : T, isLocal: Boolean): Broadcast[T] = { broadcastFactory.newBroadcast[T](value_, isLocal, nextBroadcastId.getAndIncrement()) } def unbroadcast(id: Long, removeFromDriver: Boolean, blocking: Boolean) { broadcastFactory.unbroadcast(id, removeFromDriver, blocking) } } Broadcast源码 /* * Licensed to the Apache Software Foundation (ASF) under one or more * contributor license agreements. See the NOTICE file distributed with * this work for additional information regarding copyright ownership. * The ASF licenses this file to You under the Apache License, Version 2.0 * (the "License"); you may not use this file except in compliance with * the License. You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ package org.apache.spark.broadcast import java.io.Serializable import org.apache.spark.SparkException import org.apache.spark.Logging import org.apache.spark.util.Utils import scala.reflect.ClassTag /** * A broadcast variable. Broadcast variables allow the programmer to keep a read-only variable * cached on each machine rather than shipping a copy of it with tasks. They can be used, for * example, to give every node a copy of a large input dataset in an efficient manner. Spark also * attempts to distribute broadcast variables using efficient broadcast algorithms to reduce * communication cost. * * Broadcast variables are created from a variable `v` by calling * [[org.apache.spark.SparkContext#broadcast]]. * The broadcast variable is a wrapper around `v`, and its value can be accessed by calling the * `value` method. The interpreter session below shows this: * * {{{ * scala> val broadcastVar = sc.broadcast(Array(1, 2, 3)) * broadcastVar: org.apache.spark.broadcast.Broadcast[Array[Int]] = Broadcast(0) * * scala> broadcastVar.value * res0: Array[Int] = Array(1, 2, 3) * }}} * * After the broadcast variable is created, it should be used instead of the value `v` in any * functions run on the cluster so that `v` is not shipped to the nodes more than once. * In addition, the object `v` should not be modified after it is broadcast in order to ensure * that all nodes get the same value of the broadcast variable (e.g. if the variable is shipped * to a new node later). * * @param id A unique identifier for the broadcast variable. * @tparam T Type of the data contained in the broadcast variable. */ abstract class Broadcast[T: ClassTag](val id: Long) extends Serializable with Logging { /** * Flag signifying whether the broadcast variable is valid * (that is, not already destroyed) or not. */ @volatile private var _isValid = true private var _destroySite = "" /** Get the broadcasted value. */ def value: T = { assertValid() getValue() } /** * Asynchronously delete cached copies of this broadcast on the executors. * If the broadcast is used after this is called, it will need to be re-sent to each executor. */ def unpersist() { unpersist(blocking = false) } /** * Delete cached copies of this broadcast on the executors. If the broadcast is used after * this is called, it will need to be re-sent to each executor. * @param blocking Whether to block until unpersisting has completed */ def unpersist(blocking: Boolean) { assertValid() doUnpersist(blocking) } /** * Destroy all data and metadata related to this broadcast variable. Use this with caution; * once a broadcast variable has been destroyed, it cannot be used again. * This method blocks until destroy has completed */ def destroy() { destroy(blocking = true) } /** * Destroy all data and metadata related to this broadcast variable. Use this with caution; * once a broadcast variable has been destroyed, it cannot be used again. * @param blocking Whether to block until destroy has completed */ private[spark] def destroy(blocking: Boolean) { assertValid() _isValid = false _destroySite = Utils.getCallSite().shortForm logInfo("Destroying %s (from %s)".format(toString, _destroySite)) doDestroy(blocking) } /** * Whether this Broadcast is actually usable. This should be false once persisted state is * removed from the driver. */ private[spark] def isValid: Boolean = { _isValid } /** * Actually get the broadcasted value. Concrete implementations of Broadcast class must * define their own way to get the value. */ protected def getValue(): T /** * Actually unpersist the broadcasted value on the executors. Concrete implementations of * Broadcast class must define their own logic to unpersist their own data. */ protected def doUnpersist(blocking: Boolean) /** * Actually destroy all data and metadata related to this broadcast variable. * Implementation of Broadcast class must define their own logic to destroy their own * state. */ protected def doDestroy(blocking: Boolean) /** Check if this broadcast is valid. If not valid, exception is thrown. */ protected def assertValid() { if (!_isValid) { throw new SparkException( "Attempted to use %s after it was destroyed (%s) ".format(toString, _destroySite)) } } override def toString: String = "Broadcast(" + id + ")" } 其他的，不一一赘述了。 3、累加器 为什么需要，累加器？ 答：第一种情况，是，test把数据副本运行起来。 第二种情况，有全局变量和局部变量，有了广播，为什么还需要累加器？ （3）累加器（获取全局唯一的状态对象，SparkContext创建，被Driver控制，在Text实际运行的时候，每次都可以保证修改之后获取全局唯一的对象，Driver中可读，Executor可读） 累加器是仅仅被相关操作累加的变量，因此可以在并行中被有效地支持。它可以被用来实现计数器和总和。Spark原生地只支持数字类型的累加器，编程者可以添加新类型的支持。如果创建累加器时指定了名字，可以在Spark的UI界面看到。这有利于理解每个执行阶段的进程。（对于python还不支持） 累加器通过对一个初始化了的变量v调用SparkContext.accumulator(v)来创建。在集群上运行的任务可以通过add或者"+="方法在累加器上进行累加操作。但是，它们不能读取它的值。只有驱动程序能够读取它的值，通过累加器的value方法。 累加器的特征：全局的，Accumulator：对于Executor只能修改但不可读，只对Driver可读（因为通过Driver控制整个集群的状态），不同的executor 修改，不会彼此覆盖（枷锁机制） 累加器实战： scala>val sum = sc.accumulator(0) sum: org.apache.spark.Accumulator[Int] = 0 scala>val data = sc.parallelize(1 to 100) data: org.apache.spark.rdd.RDD[Int] = ParallelCollectionRDD[63] at parallelize at <console>:21 scala>val result = data.foreach(item =>sum += item) took 6.548568 s result: Unit = () scala> println(sum) 5050 累加器在记录集群全局唯一的状态的时候极其重要，保持唯一的全局状态的变量，所以其重要性不言而喻。 Driver中取值，Executor中计算， 1、累计器全局（全集群）唯一，只增不减（Executor中的task去修改，即累加）； 2、累加器是Executor共享； 我的理解应该是对的，集群全局变量，谁操作，从driver上拿去操作，然后下个Executor在用的时候，拿上个Executor执行的结果，也就是从Driver那里拿。 accumulator源码 /* * Licensed to the Apache Software Foundation (ASF) under one or more * contributor license agreements. See the NOTICE file distributed with * this work for additional information regarding copyright ownership. * The ASF licenses this file to You under the Apache License, Version 2.0 * (the "License"); you may not use this file except in compliance with * the License. You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ package org.apache.spark import java.io.{ObjectInputStream, Serializable} import scala.collection.generic.Growable import scala.collection.Map import scala.collection.mutable import scala.ref.WeakReference import scala.reflect.ClassTag import org.apache.spark.serializer.JavaSerializer import org.apache.spark.util.Utils /** * A data type that can be accumulated, ie has an commutative and associative "add" operation, * but where the result type, `R`, may be different from the element type being added, `T`. * * You must define how to add data, and how to merge two of these together. For some data types, * such as a counter, these might be the same operation. In that case, you can use the simpler * [[org.apache.spark.Accumulator]]. They won't always be the same, though -- e.g., imagine you are * accumulating a set. You will add items to the set, and you will union two sets together. * * @param initialValue initial value of accumulator * @param param helper object defining how to add elements of type `R` and `T` * @param name human-readable name for use in Spark's web UI * @param internal if this [[Accumulable]] is internal. Internal [[Accumulable]]s will be reported * to the driver via heartbeats. For internal [[Accumulable]]s, `R` must be * thread safe so that they can be reported correctly. * @tparam R the full accumulated data (result type) * @tparam T partial data that can be added in */ class Accumulable[R, T] private[spark] ( @transient initialValue: R, param: AccumulableParam[R, T], val name: Option[String], internal: Boolean) extends Serializable { private[spark] def this( @transient initialValue: R, param: AccumulableParam[R, T], internal: Boolean) = { this(initialValue, param, None, internal) } def this(@transient initialValue: R, param: AccumulableParam[R, T], name: Option[String]) = this(initialValue, param, name, false) def this(@transient initialValue: R, param: AccumulableParam[R, T]) = this(initialValue, param, None) val id: Long = Accumulators.newId @volatile @transient private var value_ : R = initialValue // Current value on master val zero = param.zero(initialValue) // Zero value to be passed to workers private var deserialized = false Accumulators.register(this) /** * If this [[Accumulable]] is internal. Internal [[Accumulable]]s will be reported to the driver * via heartbeats. For internal [[Accumulable]]s, `R` must be thread safe so that they can be * reported correctly. */ private[spark] def isInternal: Boolean = internal /** * Add more data to this accumulator / accumulable * @param term the data to add */ def += (term: T) { value_ = param.addAccumulator(value_, term) } /** * Add more data to this accumulator / accumulable * @param term the data to add */ def add(term: T) { value_ = param.addAccumulator(value_, term) } /** * Merge two accumulable objects together * * Normally, a user will not want to use this version, but will instead call `+=`. * @param term the other `R` that will get merged with this */ def ++= (term: R) { value_ = param.addInPlace(value_, term)} /** * Merge two accumulable objects together * * Normally, a user will not want to use this version, but will instead call `add`. * @param term the other `R` that will get merged with this */ def merge(term: R) { value_ = param.addInPlace(value_, term)} /** * Access the accumulator's current value; only allowed on master. */ def value: R = { if (!deserialized) { value_ } else { throw new UnsupportedOperationException("Can't read accumulator value in task") } } /** * Get the current value of this accumulator from within a task. * * This is NOT the global value of the accumulator. To get the global value after a * completed operation on the dataset, call `value`. * * The typical use of this method is to directly mutate the local value, eg., to add * an element to a Set. */ def localValue: R = value_ /** * Set the accumulator's value; only allowed on master. */ def value_= (newValue: R) { if (!deserialized) { value_ = newValue } else { throw new UnsupportedOperationException("Can't assign accumulator value in task") } } /** * Set the accumulator's value; only allowed on master */ def setValue(newValue: R) { this.value = newValue } // Called by Java when deserializing an object private def readObject(in: ObjectInputStream): Unit = Utils.tryOrIOException { in.defaultReadObject() value_ = zero deserialized = true // Automatically register the accumulator when it is deserialized with the task closure. // // Note internal accumulators sent with task are deserialized before the TaskContext is created // and are registered in the TaskContext constructor. Other internal accumulators, such SQL // metrics, still need to register here. val taskContext = TaskContext.get() if (taskContext != null) { taskContext.registerAccumulator(this) } } override def toString: String = if (value_ == null) "null" else value_.toString } /** * Helper object defining how to accumulate values of a particular type. An implicit * AccumulableParam needs to be available when you create [[Accumulable]]s of a specific type. * * @tparam R the full accumulated data (result type) * @tparam T partial data that can be added in */ trait AccumulableParam[R, T] extends Serializable { /** * Add additional data to the accumulator value. Is allowed to modify and return `r` * for efficiency (to avoid allocating objects). * * @param r the current value of the accumulator * @param t the data to be added to the accumulator * @return the new value of the accumulator */ def addAccumulator(r: R, t: T): R /** * Merge two accumulated values together. Is allowed to modify and return the first value * for efficiency (to avoid allocating objects). * * @param r1 one set of accumulated data * @param r2 another set of accumulated data * @return both data sets merged together */ def addInPlace(r1: R, r2: R): R /** * Return the "zero" (identity) value for an accumulator type, given its initial value. For * example, if R was a vector of N dimensions, this would return a vector of N zeroes. */ def zero(initialValue: R): R } private[spark] class GrowableAccumulableParam[R <% Growable[T] with TraversableOnce[T] with Serializable: ClassTag, T] extends AccumulableParam[R, T] { def addAccumulator(growable: R, elem: T): R = { growable += elem growable } def addInPlace(t1: R, t2: R): R = { t1 ++= t2 t1 } def zero(initialValue: R): R = { // We need to clone initialValue, but it's hard to specify that R should also be Cloneable. // Instead we'll serialize it to a buffer and load it back. val ser = new JavaSerializer(new SparkConf(false)).newInstance() val copy = ser.deserialize[R](ser.serialize(initialValue)) copy.clear() // In case it contained stuff copy } } /** * A simpler value of [[Accumulable]] where the result type being accumulated is the same * as the types of elements being merged, i.e. variables that are only "added" to through an * associative operation and can therefore be efficiently supported in parallel. They can be used * to implement counters (as in MapReduce) or sums. Spark natively supports accumulators of numeric * value types, and programmers can add support for new types. * * An accumulator is created from an initial value `v` by calling [[SparkContext#accumulator]]. * Tasks running on the cluster can then add to it using the [[Accumulable#+=]] operator. * However, they cannot read its value. Only the driver program can read the accumulator's value, * using its value method. * * The interpreter session below shows an accumulator being used to add up the elements of an array: * * {{{ * scala> val accum = sc.accumulator(0) * accum: spark.Accumulator[Int] = 0 * * scala> sc.parallelize(Array(1, 2, 3, 4)).foreach(x => accum += x) * ... * 10/09/29 18:41:08 INFO SparkContext: Tasks finished in 0.317106 s * * scala> accum.value * res2: Int = 10 * }}} * * @param initialValue initial value of accumulator * @param param helper object defining how to add elements of type `T` * @tparam T result type */ class Accumulator[T] private[spark] ( @transient private[spark] val initialValue: T, param: AccumulatorParam[T], name: Option[String], internal: Boolean) extends Accumulable[T, T](initialValue, param, name, internal) { def this(initialValue: T, param: AccumulatorParam[T], name: Option[String]) = { this(initialValue, param, name, false) } def this(initialValue: T, param: AccumulatorParam[T]) = { this(initialValue, param, None, false) } } /** * A simpler version of [[org.apache.spark.AccumulableParam]] where the only data type you can add * in is the same type as the accumulated value. An implicit AccumulatorParam object needs to be * available when you create Accumulators of a specific type. * * @tparam T type of value to accumulate */ trait AccumulatorParam[T] extends AccumulableParam[T, T] { def addAccumulator(t1: T, t2: T): T = { addInPlace(t1, t2) } } object AccumulatorParam { // The following implicit objects were in SparkContext before 1.2 and users had to // `import SparkContext._` to enable them. Now we move them here to make the compiler find // them automatically. However, as there are duplicate codes in SparkContext for backward // compatibility, please update them accordingly if you modify the following implicit objects. implicit object DoubleAccumulatorParam extends AccumulatorParam[Double] { def addInPlace(t1: Double, t2: Double): Double = t1 + t2 def zero(initialValue: Double): Double = 0.0 } implicit object IntAccumulatorParam extends AccumulatorParam[Int] { def addInPlace(t1: Int, t2: Int): Int = t1 + t2 def zero(initialValue: Int): Int = 0 } implicit object LongAccumulatorParam extends AccumulatorParam[Long] { def addInPlace(t1: Long, t2: Long): Long = t1 + t2 def zero(initialValue: Long): Long = 0L } implicit object FloatAccumulatorParam extends AccumulatorParam[Float] { def addInPlace(t1: Float, t2: Float): Float = t1 + t2 def zero(initialValue: Float): Float = 0f } // TODO: Add AccumulatorParams for other types, e.g. lists and strings } // TODO: The multi-thread support in accumulators is kind of lame; check // if there's a more intuitive way of doing it right private[spark] object Accumulators extends Logging { /** * This global map holds the original accumulator objects that are created on the driver. * It keeps weak references to these objects so that accumulators can be garbage-collected * once the RDDs and user-code that reference them are cleaned up. */ val originals = mutable.Map[Long, WeakReference[Accumulable[_, _]]]() private var lastId: Long = 0 def newId(): Long = synchronized { lastId += 1 lastId } def register(a: Accumulable[_, _]): Unit = synchronized { originals(a.id) = new WeakReference[Accumulable[_, _]](a) } def remove(accId: Long) { synchronized { originals.remove(accId) } } // Add values to the original accumulators with some given IDs def add(values: Map[Long, Any]): Unit = synchronized { for ((id, value) <- values) { if (originals.contains(id)) { // Since we are now storing weak references, we must check whether the underlying data // is valid. originals(id).get match { case Some(accum) => accum.asInstanceOf[Accumulable[Any, Any]] ++= value case None => throw new IllegalAccessError("Attempted to access garbage collected Accumulator.") } } else { logWarning(s"Ignoring accumulator update for unknown accumulator id $id") } } } } private[spark] object InternalAccumulator { val PEAK_EXECUTION_MEMORY = "peakExecutionMemory" val TEST_ACCUMULATOR = "testAccumulator" // For testing only. // This needs to be a def since we don't want to reuse the same accumulator across stages. private def maybeTestAccumulator: Option[Accumulator[Long]] = { if (sys.props.contains("spark.testing")) { Some(new Accumulator( 0L, AccumulatorParam.LongAccumulatorParam, Some(TEST_ACCUMULATOR), internal = true)) } else { None } } /** * Accumulators for tracking internal metrics. * * These accumulators are created with the stage such that all tasks in the stage will * add to the same set of accumulators. We do this to report the distribution of accumulator * values across all tasks within each stage. */ def create(sc: SparkContext): Seq[Accumulator[Long]] = { val internalAccumulators = Seq( // Execution memory refers to the memory used by internal data structures created // during shuffles, aggregations and joins. The value of this accumulator should be // approximately the sum of the peak sizes across all such data structures created // in this task. For SQL jobs, this only tracks all unsafe operators and ExternalSort. new Accumulator( 0L, AccumulatorParam.LongAccumulatorParam, Some(PEAK_EXECUTION_MEMORY), internal = true) ) ++ maybeTestAccumulator.toSeq internalAccumulators.foreach { accumulator => sc.cleaner.foreach(_.registerAccumulatorForCleanup(accumulator)) } internalAccumulators } } 本文转自大数据躺过的坑博客园博客，原文链接：http://www.cnblogs.com/zlslch/p/5914696.html，如需转载请自行联系原作者
