當前位置：首頁 > 编程资源 > 编程问答 >内容正文

编程问答

Spark RDD概念学习系列之rdd持久化、广播、累加器（十八）

發布時間：2025/4/16 编程问答 22 豆豆

生活随笔收集整理的這篇文章主要介紹了 Spark RDD概念学习系列之rdd持久化、广播、累加器（十八）小編覺得挺不錯的,現在分享給大家,幫大家做個參考.

1、rdd持久化

2、廣播

3、累加器

1、rdd持久化

　　通過spark-shell，可以快速的驗證我們的想法和操作！

啟動hdfs集群

spark@SparkSingleNode:/usr/local/hadoop/hadoop-2.6.0$ sbin/start-dfs.sh

啟動spark集群

spark@SparkSingleNode:/usr/local/spark/spark-1.5.2-bin-hadoop2.6$ sbin/start-all.sh

?啟動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

可見，reduce是個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源碼

/**
* 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"))
} 可見,這也是一個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源碼

/**
* 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: _*)
} 可見,這也是一個action操作。

　　從收集結果的角度來說，如果想要在命令行終端中，看到執行結果，就必須collect。

?　　從源碼的角度來說，凡是action級別的操作，都會觸發sc.rubJob。這點，spark里是一個應用程序允許有多個Job,而hadoop里一個應用程序只能一個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源碼?

/**
* Return the number of elements in the RDD.
*/
def count(): Long = sc.runJob(this, Utils.getIteratorSize _).sum

可見,這也是一個action操作。

take

scala> val topN = numbers.take(5)

topN: Array[Int] = Array(1, 2, 3, 4, 5)

take源碼

/**
* 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
}
} 可見,這也是一個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源碼

/**
* 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
}

可見,這也是一個action操作。

saveAsTextFile
之前，在 rdd實戰（rdd基本操作實戰及transformation和action流程圖）（源碼） scala>?val partitionsReadmeRdd = ?sc.textFile("hdfs://SparkSingleNode:9000/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).saveAsTextFile("~/partition1README.txt")

這里呢。

scala>?val partitionsReadmeRdd = ?sc.textFile("/README.md").flatMap(_.split(" ")).map(word =>(word,1)).reduceByKey(_+_,1).saveAsTextFile("/partition1README.txt")

saveAsTextFile源碼

/**
* 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不僅，可保存在集群里，也可以保存到本地，這就要看hadoop的運行模式。
由此可見，它也是個action操作。
以上是rdd持久化的第一個方面，就是action級別的操作。
rdd持久化的第二個方面，就是通過persist。
為什么在spark里，隨處可見persist的身影呢？
原因一：spark在默認情況下，數據是放在內存中，適合高速迭代。比如在一個stage里，有1000個步驟，它其實只在第1個步驟輸入數據，在第1000個步驟輸出數據，在中間不產生臨時數據。但是，分布式系統，分享非常高，所以，容出錯，設計到容錯。

　　　　由于，rdd是有血統繼承關系的，即lineager。如果后面的rdd數據分片出錯了或rdd本身出錯了，則，可根據其前面依賴的lineager，算出來。
但是，假設1000個步驟，如果之前，沒有父rdd進行persist或cache的話，則要重頭開始了。親！

什么時候，該persist？ 1、在某個步驟非常費時的情況下，不好使　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　（手動）
2、計算鏈條特別長的情況下　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　（手動） 3、checkpoint所在的rdd也一定要持久化數據（注意：在checkpoint之前，進行persist） 　　　　　　（手動） checkpoint是rdd的算子， 　　先寫，某個具體rdd.checkpoint 或某個具體rdd.cache ，再寫，某個具體rdd.persist 4、shuffle之后（因為shuffle之后，要網絡傳輸，風險大） 　　　　　　　　　　　　　　　　　　　　　　　　（手動） 5、shuffle之前（框架，默認給我們做的，把數據持久化到本地磁盤）

checkpoint源碼

/**
* 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源碼
/**
* 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里有很多類型

這里，牽扯到序列化。

問，為什么要序列化？

答：節省空間，減少體積。內存不夠時，把MEMORY中的數據，進行序列化。

　　當然，也有不好一面，序列化時，會反序列化，反序列化耗cpu。

MEMORY_AND_DISK
假設，我們制定數據存儲方式是，MEMORY_AND_DISK。則，是不是同時，存儲到內存和磁盤呢？
答：不是啊，親。spark一定是優先考慮內存的啊，只要內存足夠，不會考慮磁盤。若內存不夠了，則才放部分數據到磁盤。

極大地減少數據丟失概率發生。

MEMORY_ONLY 假設，我們制定數據存儲方式是，MEMORY_ONLY。則，只放到內存。當內存不夠了，會出現OOM?；驍祿G失。

OFF_HEAP 這牽扯到Tachyon，基于內存的分布式系統

為什么有2分副本？好處是？

假設，一個計算特別耗時，而且，又是基于內存，如果其中一份副本崩潰掉，則可迅速切換到另一份副本去計算。這就是“空間換時間”！非常重要
這不是并行計算，這是計算后的結果，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！

實際工程中， cache之后，如果有其他算子，則會，重新觸發這個工作過程。

注意:cache，不是action

?cache緩存，怎么讓它失效？

答：unpersist

persist是lazy級別的，unpersist是eager級別的。cache是persist的一個特殊情況。

cache和persist的區別？

答：persist可以放到磁盤、放到內存、同時放到內存和磁盤。以及多份副本

　　cache只能放到內存，以及只能一份副本。

persisit在內存不夠時，保存在磁盤的哪個目錄？

答：local的process。

好的，以上是，rdd持久化的兩個方面。

rdd持久化的第一個方面，就是常用的action級別的操作。 rdd持久化的第二個方面，就是持久化的不同方式，以及它內部的運行情況

　　小知識：cache之后，一定不能立即有其他算子！實際工程中， cache之后，如果有其他算子，則會，重新觸發這個工作過程。

?一般都不會跨機器抓內存，寧愿排隊。寧愿數據不動代碼動。

2、廣播

?　　為什么要有，rdd廣播？

　　答：大變量、join、冗余、減少數據移動、通信、狀態、集群消息、共享、網絡傳輸慢要提前、數據量大耗網絡、減少通信、要同步。

?　　為什么大變量，需要廣播呢？

?　　答：原因是，每個task運行，讀取全集數據時，task本身執行時，每次都要拷貝一份數據副本，如果變量比較大，如一百萬，則要拷貝一百萬。

（2）廣播（線程中共享）不必每個task都拷貝一份副本，因為它是全局唯一的，極大的減少oom，減少通信，冗余、共享變量等。廣播是將數據廣播到Executor的內存中，其內部所以的任務都會只享有全局唯一的變量，減少網絡傳輸。 text在讀取數據時候，拷貝一份的數據副本（變量），因為函數式編程（變量不變），不拷貝狀態容易被改變，數據量小（1、引用較小2、數據本身小），變量大容易產生oom（task拷貝數據在內存中運行），網絡傳輸慢，要提前，冗余、共享，減少通信。 廣播變量：

廣播變量允許程序員將一個只讀的變量緩存在每臺機器上，而不用在任務之間傳遞變量。廣播變量可被用于有效地給每個節點一個大輸入數據集的副本。Spark還嘗試使用高效地廣播算法來分發變量，進而減少通信的開銷。

Spark的動作通過一系列的步驟執行，這些步驟由分布式的洗牌操作分開。Spark自動地廣播每個步驟每個任務需要的通用數據。這些廣播數據被序列化地緩存，在運行任務之前被反序列化出來。這意味著當我們需要在多個階段的任務之間使用相同的數據，或者以反序列化形式緩存數據是十分重要的時候，顯式地創建廣播變量才有用。

（本段摘自：http://blog.csdn.net/happyanger6/article/details/46576831）

　　　　　　　　　　　　　　　　　　　　　　　　　　　　廣播工作機制圖?

?　　　　　　　　　　　　　　　　　　　　　　　　　　　　廣播工作機制圖?

參考：　http://blog.csdn.net/kxr0502/article/details/50574561

問：讀廣播，會消耗網絡傳輸嗎？

答：不消耗，廣播是放在內存中。讀取它，不消耗。

問：廣播變量是不是就是向每一個executor，廣播一份數據，而不是向每一個task，廣播一份數據？這樣對嗎?

答：對

　　廣播是由Driver發給當前Application分配的所有Executor內存級別的全局只讀變量，Executor中的線程池中的線程共享該全局變量，極大的減少了網絡傳輸（否則的話每個Task都要傳輸一次該變量）并極大的節省了內存，當然也隱形的提高的CPU的有效工作。

?實戰創建廣播：

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>:23
broadcastNumber: 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>

　　我們知道，是test是要廣播變量，但，我們編程，對rdd。

//通過在一個變量v上調用SparkContext.broadcast(v)可以創建廣播變量。廣播變量是圍繞著v的封裝，可以通過value方法訪問這個變量。

問：廣播變量里有很多變量嗎？

答:當然可以有很多，用java bin或scala封裝，就可以了。

　　如，在這里。廣播變量是，broadcastNumber，里，有變量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，可見，通過機制、流程圖和實戰，深度剖析對廣播全面詳解！

broadcast源碼分析

?參考：?http://www.cnblogs.com/seaspring/p/5682053.html

BroadcastManager源碼

Broadcast源碼

?其他的，不一一贅述了。

?3、累加器

?　　為什么需要，累加器？

　　答：第一種情況，是，test把數據副本運行起來。

　　 ? ? ? ?第二種情況，有全局變量和局部變量，有了廣播，為什么還需要累加器？

（3）累加器（獲取全局唯一的狀態對象，SparkContext創建，被Driver控制，在Text實際運行的時候，每次都可以保證修改之后獲取全局唯一的對象，Driver中可讀，Executor可讀）

? ??? ??累加器是僅僅被相關操作累加的變量，因此可以在并行中被有效地支持。它可以被用來實現計數器和總和。Spark原生地只支持數字類型的累加器，編程者可以添加新類型的支持。如果創建累加器時指定了名字，可以在Spark的UI界面看到。這有利于理解每個執行階段的進程。（對于python還不支持）

? ??? ??累加器通過對一個初始化了的變量v調用SparkContext.accumulator(v)來創建。在集群上運行的任務可以通過add或者"+="方法在累加器上進行累加操作。但是，它們不能讀取它的值。只有驅動程序能夠讀取它的值，通過累加器的value方法。

?　　累加器的特征：全局的，Accumulator：對于Executor只能修改但不可讀，只對Driver可讀（因為通過Driver控制整個集群的狀態），不同的executor 修改，不會彼此覆蓋（枷鎖機制）

?累加器實戰：

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?在記錄集群全局唯一的狀態的時候極其重要，保持唯一的全局狀態的變量，所以其重要性不言而喻。 Driver中取值，Executor中計算， 1、累計器全局（全集群）唯一，只增不減（Executor中的task去修改，即累加）； 2、累加器是Executor共享；我的理解應該是對的，集群全局變量，誰操作，從driver上拿去操作，然后下個Executor在用的時候，拿上個Executor執行的結果，也就是從Driver那里拿。

?accumulator源碼

/*
* 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
}
}

參考

王家林老師是大數據技術集大成者，中國Spark第一人：

ＤＴ大數據夢工廠

新浪微博：www.weibo.com/ilovepains/

微信公眾號：DT_Spark

博客：http://.blog.sina.com.cn/ilovepains

TEL:18610086859

Email:18610086859@vip.126.com

參考鏈接：

http://blog.csdn.net/kxr0502/article/details/50574561

http://blog.csdn.net/happyanger6/article/details/46576831

http://blog.csdn.net/happyanger6/article/details/46552823

《新程序員》：云原生和全面數字化實踐50位技術專家共同創作，文字、視頻、音頻交互閱讀

總結

以上是生活随笔為你收集整理的Spark RDD概念学习系列之rdd持久化、广播、累加器（十八）的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。

上一篇：论文笔记之：Multiple Featu
下一篇：用YII实现多重查询(基于tag)