问题导读:
1. 排序算子是如何做排序的?
2. 完整的排序流程是?
解决方案:
1 前言
在前面一系列博客中,特别在Shuffle博客系列中,曾描述过在生成ShuffleWrite的文件的时候,对每个partition会先进行排序并spill到文件中,最后合并成ShuffleWrite的文件,也就是每个Partition里的内容已经进行了排序,在最后的action操作的时候需要对每个executor生成的shuffle文件相同的Partition进行合并,完成Action的操作。
排序算子和常见的reduce算子算法有何区别?
常见的一些聚合、reduce算子,不需要排序
- 将相同的hashcode分配到同一个partition,哪怕是不同的executor
- 在做最后的合并的时候,只需要合并不同的executor里相同的partition就可以了
- 对每个partition进行排序,考虑内存因数,解决相同的Partition多文件合并的问题,使用外排序进行相同的key合并
2 排序
下面是一个常见的排序的小例子:
[mw_shl_code=scala,true]
package spark.sort
import org.apache.spark.SparkConf
import org.apache.spark.SparkContext
object sortsample {
def main(args: Array[String]) {
val conf = new SparkConf().setAppName("sortsample")
val sc = new SparkContext(conf)
var pairs = sc.parallelize(Array(("a",0),("b",0),("c",3),("d",6),("e",0),("f",0),("g",3),("h",6)), 2);
pairs.sortByKey(true, 3).collect().foreach(println);
}
}[/mw_shl_code]
核心代码:OrderedRDDFunctions.scala
会很奇怪么?RDD里面并没有sortByKey的方法?在这里和前面博客里提到的PairRDDFunctions一样,隐式转换:
[mw_shl_code=scala,true] implicit def rddToOrderedRDDFunctions[K : Ordering : ClassTag, V: ClassTag](rdd: RDD[(K, V)])
: OrderedRDDFunctions[K, V, (K, V)] = {
new OrderedRDDFunctions[K, V, (K, V)](rdd)
}[/mw_shl_code]
调用的是OrderedRDDFunctions.scala里的方法
[mw_shl_code=scala,true] def sortByKey(ascending: Boolean = true, numPartitions: Int = self.partitions.length)
: RDD[(K, V)] = self.withScope
{
val part = new RangePartitioner(numPartitions, self, ascending)
new ShuffledRDD[K, V, V](self, part)
.setKeyOrdering(if (ascending) ordering else ordering.reverse)
}[/mw_shl_code]
对Partition采用了范围分配的策略,为何要使用范围分配的策略?
- 对其它非排序类型的算子,使用散列算法,只要保证相同的key是分配在相同的partition就可以了,并不会影响相同的key的合并,计算。
- 对排序来说,如果只是保证相同的key在相同的Partition并不足够,最后还是需要合并所有的Partition进行排序合并,如果这发生在Driver端做这件事,将会非常可怕,那么我们可以做一些策略改变,制定一些Range,使排序相近的key分配到同一个Range上,在把Range扩大化,比如:一个Partition管理一个Range
2.1 分配Range
Range的分配不合理,会影响数据的不均衡,导致executor在做同Partition排序的时候会不均衡,并行计算的整体性能往往会被单个最糟糕的运行节点所拖累,如果提高运算的速度,需要考虑数据分配的均衡性。
2.1.1 每个区块采样大小
获取所有的key,依据所有的Key制定区间,这显然是不明智的,后果变成一个全量数据的排序。我们可以采用部分采样的策略,基于采样数据进行区间划分,首先我们需要评估一个简单的采样大小的阈值。
Partitioner.scala rangeBounds
代码如下:
[mw_shl_code=scala,true]val sampleSize = math.min(20.0 * partitions, 1e6)
// Assume the input partitions are roughly balanced and over-sample a little bit.
val sampleSizePerPartition = math.ceil(3.0 * sampleSize / rdd.partitions.length).toInt
val (numItems, sketched) = RangePartitioner.sketch(rdd.map(_._1), sampleSizePerPartition)[/mw_shl_code]
partitions: 参数在指定sortByKey的时候设置的区块大小:3
[mw_shl_code=scala,true]pairs.sortByKey(true, 3) [/mw_shl_code]
rdd.partitions: 指的是在数据的分区块大小:2
[mw_shl_code=scala,true]sc.parallelize(Array(("a",0),("b",0),("c",3),("d",6),("e",0),("f",0),("g",3),("h",6)), 2)[/mw_shl_code]
每个区块需要采样的数量是通过几个固定参数来计算
[mw_shl_code=scala,true]val sampleSizePerPartition = math.ceil(3.0 * sampleSize / rdd.partitions.length).toInt[/mw_shl_code]
2.1.2 Sketch采样(蓄水池采样法)
[mw_shl_code=scala,true] def sketch[K : ClassTag](
rdd: RDD[K],
sampleSizePerPartition: Int): (Long, Array[(Int, Long, Array[K])]) = {
val shift = rdd.id
// val classTagK = classTag[K] // to avoid serializing the entire partitioner object
val sketched = rdd.mapPartitionsWithIndex { (idx, iter) =>
val seed = byteswap32(idx ^ (shift << 16))
val (sample, n) = SamplingUtils.reservoirSampleAndCount(
iter, sampleSizePerPartition, seed)
Iterator((idx, n, sample))
}.collect()
val numItems = sketched.map(_._2).sum
(numItems, sketched)
}[/mw_shl_code]
mapPartitionsWithIndex, collection 这些都是RDD ,都是需要在提交job进行运算的,也就是采样的过程中,是通过executor执行了一次job
[mw_shl_code=scala,true] def reservoirSampleAndCount[T: ClassTag](
input: Iterator[T],
k: Int,
seed: Long = Random.nextLong())
: (Array[T], Long) = {
val reservoir = new Array[T](k)
// Put the first k elements in the reservoir.
var i = 0
while (i < k && input.hasNext) {
val item = input.next()
reservoir(i) = item
i += 1
}
// If we have consumed all the elements, return them. Otherwise do the replacement.
if (i < k) {
// If input size < k, trim the array to return only an array of input size.
val trimReservoir = new Array[T](i)
System.arraycopy(reservoir, 0, trimReservoir, 0, i)
(trimReservoir, i)
} else {
// If input size > k, continue the sampling process.
var l = i.toLong
val rand = new XORShiftRandom(seed)
while (input.hasNext) {
val item = input.next()
l += 1
// There are k elements in the reservoir, and the l-th element has been
// consumed. It should be chosen with probability k/l. The expression
// below is a random long chosen uniformly from [0,l)
val replacementIndex = (rand.nextDouble() * l).toLong
if (replacementIndex < k) {
reservoir(replacementIndex.toInt) = item
}
}
(reservoir, l)
}
}[/mw_shl_code]
函数reservoirSampleAndCount采样
- 当数据小于要采样的集合的时候,可以使用数据为样本
- 当数据集合超过需要采样数目的时候会继续遍历整个数据集合,通过随机数进行位置的随机替换,保证采样数据的随机性
返回的结果里包含了总数据集,区块编号,区块的数量,每个区块的采样集
2.1.3 重新采样
为了避免某些区块的数据量过大,设置了一个阈值:
[mw_shl_code=scala,true]val fraction = math.min(sampleSize / math.max(numItems, 1L), 1.0) [/mw_shl_code]
阈值=采样数除于总数据量,当某个区块的数据量*阈值大于每个区的采样率的时候,认为这个区块的采样率是不足的,需要重新采样
[mw_shl_code=scala,true]val imbalanced = new PartitionPruningRDD(rdd.map(_._1), imbalancedPartitions.contains)
val seed = byteswap32(-rdd.id - 1)
val reSampled = imbalanced.sample(withReplacement = false, fraction, seed).collect()
val weight = (1.0 / fraction).toFloat
candidates ++= reSampled.map(x => (x, weight))[/mw_shl_code]
2.1.4 采样集key的权重
我们在前面对每个区进行了相同数量的采样(不包含重新采样),但是每个区的数量有可能是不均衡的,为了避免不均衡性需要对每个区采样的key进行权重设置,尽量分配高权重给数据量多的区
权重因子:
[mw_shl_code=scala,true]val weight = (n.toDouble / sample.length).toFloat[/mw_shl_code]
n 是区的数据数量
sample 是采样的数量
这里权重的最小值是1,因为采样的数量肯定是小于等于数据
当数据量大于采样数量的时候,每个区的采样数量是相同的,那么意味着区的数据量越大,该区块的key的权重也就越大
2.1.5 分配每个区块的range
样本已经采集好了,现在需要对依据样本进行区块的range进行分配
- 先对样本进行排序
- 依据每个样本的权重计算每个区块平均所分配的权重
- 最后通过每个区分配的权重按照顺序来决定获取哪些样本用作range,一个区分配一个样本区间
[mw_shl_code=scala,true] def determineBounds[K : Ordering : ClassTag](
candidates: ArrayBuffer[(K, Float)],
partitions: Int): Array[K] = {
val ordering = implicitly[Ordering[K]]
val ordered = candidates.sortBy(_._1)
val numCandidates = ordered.size
val sumWeights = ordered.map(_._2.toDouble).sum
val step = sumWeights / partitions
var cumWeight = 0.0
var target = step
val bounds = ArrayBuffer.empty[K]
var i = 0
var j = 0
var previousBound = Option.empty[K]
while ((i < numCandidates) && (j < partitions - 1)) {
val (key, weight) = ordered(i)
cumWeight += weight
if (cumWeight >= target) {
// Skip duplicate values.
if (previousBound.isEmpty || ordering.gt(key, previousBound.get)) {
bounds += key
target += step
j += 1
previousBound = Some(key)
}
}
i += 1
}
bounds.toArray
}[/mw_shl_code]
2.2 ShuffleWriter
在以前的博客里介绍了SortShuffleWrite,在sortByKey的排序情况下使用了BypassMergeSortShuffleWriter,把焦点聚焦到key如何分配到Partitioner和每个Partition的文件将会如何写入key,value生成Shuffle文件,在这两点上BypassMergeSortShuffleWriter将明显的不同于SortShuffleWrite
[mw_shl_code=scala,true]while (records.hasNext()) {
final Product2<K, V> record = records.next();
final K key = record._1();
partitionWriters[partitioner.getPartition(key)].write(key, record._2());
}[/mw_shl_code]
2.2.1 分配key到Partition
在函数调用了partitioner.getPartition方法,还是回到RangePartitioner类中
[mw_shl_code=scala,true] def getPartition(key: Any): Int = {
val k = key.asInstanceOf[K]
var partition = 0
if (rangeBounds.length <= 128) {
// If we have less than 128 partitions naive search
while (partition < rangeBounds.length && ordering.gt(k, rangeBounds(partition))) {
partition += 1
}
} else {
// Determine which binary search method to use only once.
partition = binarySearch(rangeBounds, k)
// binarySearch either returns the match location or -[insertion point]-1
if (partition < 0) {
partition = -partition-1
}
if (partition > rangeBounds.length) {
partition = rangeBounds.length
}
}
if (ascending) {
partition
} else {
rangeBounds.length - partition
}
}
[/mw_shl_code]
- 当Partition的分配数小于128的时候,轮训的查找每个Partition
- 当Partition大于128的时候,使用二分法查找Partition
2.2.2 生成shuffle文件
- 基于前面对key进行排序的partition的分配,写到对应的partition文件中
- 合并Partition文件生成index和data文件(shuffle_shuffleid_mapid_0.index)(shuffle_shuffleid_mapid_0.data)因为Partition已经合并了,最后一位reduceID都是为0
注意:在这里并没有象SortShuffleWrite 对每个Partition进行排序,Spill 文件,最后合并文件,而是直接写到了Partition文件中。
2.3 Shuffle Read读取Shuffle文件
在BlockStoreShuffleReader的read函数里
[mw_shl_code=scala,true]
dep.keyOrdering match {
case Some(keyOrd: Ordering[K]) =>
// Create an ExternalSorter to sort the data. Note that if spark.shuffle.spill is disabled,
// the ExternalSorter won't spill to disk.
val sorter =
new ExternalSorter[K, C, C](context, ordering = Some(keyOrd), serializer = dep.serializer)
sorter.insertAll(aggregatedIter)
context.taskMetrics().incMemoryBytesSpilled(sorter.memoryBytesSpilled)
context.taskMetrics().incDiskBytesSpilled(sorter.diskBytesSpilled)
context.taskMetrics().incPeakExecutionMemory(sorter.peakMemoryUsedBytes)
CompletionIterator[Product2[K, C], Iterator[Product2[K, C]]](sorter.iterator, sorter.stop())
case None =>
aggregatedIter
}[/mw_shl_code]
ExternalSorter.insertAll函数
[mw_shl_code=scala,true] while (records.hasNext) {
addElementsRead()
val kv = records.next()
buffer.insert(getPartition(kv._1), kv._1, kv._2.asInstanceOf[C])
maybeSpillCollection(usingMap = false)
}[/mw_shl_code]
ExternalSorter函数,这个函数在前面的这篇博客里介绍的比较清楚,这里使用了buffer结构体
[mw_shl_code=scala,true] @volatile private var map = new PartitionedAppendOnlyMap[K, C]
@volatile private var buffer = new PartitionedPairBuffer[K, C][/mw_shl_code]
在reduceByKey的这些算子相同的Key是需要合并的,所以需要使用Map结构处理相同的Key的值的合并问题,而对排序来说,并不需要相同的值合并,使用Array结构就可以了。
注:在Spark上实现Map、Array都使用了数组的结构,并没有用链表结构
在上图的PartitionPairBuffer结构中,有以下几点要注意:
插入KV结构的时候,不进行排序,也就是在处理相同的Partition的时候直接读取插入Array
会存在当内存不够Spill到磁盘的情况,关于Spill请具体参考博客链接
2.3.1 排序
当ExternalSorter.insertAll函数完成后,才会构建一个排序的迭代器
[mw_shl_code=scala,true] def partitionedIterator: Iterator[(Int, Iterator[Product2[K, C]])] = {
val collection: WritablePartitionedPairCollection[K, C] = if (usingMap) map else buffer
val usingMap = aggregator.isDefined
if (spills.isEmpty) {
// Special case: if we have only in-memory data, we don't need to merge streams, and perhaps
// we don't even need to sort by anything other than partition ID
if (!ordering.isDefined) {
// The user hasn't requested sorted keys, so only sort by partition ID, not key
groupByPartition(destructiveIterator(collection.partitionedDestructiveSortedIterator(None)))
} else {
// We do need to sort by both partition ID and key
groupByPartition(destructiveIterator(
collection.partitionedDestructiveSortedIterator(Some(keyComparator))))
}
} else {
// Merge spilled and in-memory data
merge(spills, destructiveIterator(
collection.partitionedDestructiveSortedIterator(comparator)))
}
}[/mw_shl_code]
这里分成两种情况:
还在内存里没有Spill到文件中去,这时候构建一个内存里的PartitionedDestructiveSortedIterator迭代器,在迭代器中已经排序好了PartitionPairBuffer里的内容
[mw_shl_code=scala,true] /** Iterate through the data in a given order. For this class this is not really destructive. */
override def partitionedDestructiveSortedIterator(keyComparator: Option[Comparator[K]])
: Iterator[((Int, K), V)] = {
val comparator = keyComparator.map(partitionKeyComparator).getOrElse(partitionComparator)
new Sorter(new KVArraySortDataFormat[(Int, K), AnyRef]).sort(data, 0, curSize, comparator)
iterator
}[/mw_shl_code]
Spill到文件里的,文件里的已经排好序了,需要对内存里的PartitionPairBuffer进行排序(和前面一种情况相同的处理),最后对文件和内存进行外排序(外排序可参考博客)
2.4 最后的归并
在Driver端Dag-scheduler-event-loop 线程中会处理每个executor返回的结果(刚才Partition排序后的结果)
[mw_shl_code=scala,true] private[scheduler] def handleTaskCompletion(event: CompletionEvent) {
....
case Success =>
stage.pendingPartitions -= task.partitionId
task match {
case rt: ResultTask[_, _] =>
// Cast to ResultStage here because it's part of the ResultTask
// TODO Refactor this out to a function that accepts a ResultStage
val resultStage = stage.asInstanceOf[ResultStage]
resultStage.activeJob match {
case Some(job) =>
if (!job.finished(rt.outputId)) {
updateAccumulators(event)
job.finished(rt.outputId) = true
job.numFinished += 1
// If the whole job has finished, remove it
if (job.numFinished == job.numPartitions) {
markStageAsFinished(resultStage)
cleanupStateForJobAndIndependentStages(job)
listenerBus.post(
SparkListenerJobEnd(job.jobId, clock.getTimeMillis(), JobSucceeded))
}
// taskSucceeded runs some user code that might throw an exception. Make sure
// we are resilient against that.
try {
job.listener.taskSucceeded(rt.outputId, event.result)
} catch {
case e: Exception =>
// TODO: Perhaps we want to mark the resultStage as failed?
job.listener.jobFailed(new SparkDriverExecutionException(e))
}
}
}[/mw_shl_code]
通过方法taskSucceeded的方法进行不同的Partition的合并
[mw_shl_code=scala,true]job.listener.taskSucceeded(rt.outputId, event.result)[/mw_shl_code]
[mw_shl_code=scala,true] override def taskSucceeded(index: Int, result: Any): Unit = {
// resultHandler call must be synchronized in case resultHandler itself is not thread safe.
synchronized {
resultHandler(index, result.asInstanceOf[T])
}
if (finishedTasks.incrementAndGet() == totalTasks) {
jobPromise.success(())
}
}[/mw_shl_code]
实际上是调用了resultHandler方法,我们来看看resultHandler是怎样定义的
[mw_shl_code=scala,true] def runJob[T, U: ClassTag](
rdd: RDD[T],
func: (TaskContext, Iterator[T]) => U,
partitions: Seq[Int]): Array[U] = {
val results = new Array[U](partitions.size)
runJob[T, U](rdd, func, partitions, (index, res) => results(index) = res)
results
}[/mw_shl_code]
在runJob的方法里
[mw_shl_code=scala,true] def runJob[T, U: ClassTag](
rdd: RDD[T],
func: (TaskContext, Iterator[T]) => U,
partitions: Seq[Int],
resultHandler: (Int, U) => Unit): Unit = {
if (stopped.get()) {
throw new IllegalStateException("SparkContext has been shutdown")
}
val callSite = getCallSite
val cleanedFunc = clean(func)
logInfo("Starting job: " + callSite.shortForm)
if (conf.getBoolean("spark.logLineage", false)) {
logInfo("RDD's recursive dependencies:\n" + rdd.toDebugString)
}
dagScheduler.runJob(rdd, cleanedFunc, partitions, callSite, resultHandler, localProperties.get)
progressBar.foreach(_.finishAll())
rdd.doCheckpoint()
}[/mw_shl_code]
就是:
[mw_shl_code=scala,true](index, res) => results(index) = res)[/mw_shl_code]
构建了一个数组result,将每个Partition的数值保存到result的数组里
result[0]=partition[0] =array(tuple<k,v>,tuple<k,v>.....)
什么时候对所有的Partition最后合并呢?
来看RDD的collect算子
[mw_shl_code=scala,true] def collect(): Array[T] = withScope {
val results = sc.runJob(this, (iter: Iterator[T]) => iter.toArray)
Array.concat(results: _*)
}[/mw_shl_code]
runJob返回的是result的数组,每个Partition是管理不同的范围,最后的合并只要简单的将不同的Partition合并就可以了
3. 排序完整的流程
- Driver 提交一个采样任务,需要Executor对每个Partition进行数据采样,数据采样是一次全数据的扫描
- Driver 获取采样数据,每个Partition的数据量,依据数据量的权重,进行Range的分配
- Driver 开始进行排序,先提交ShuffleMapTask ,Executor对分配到自己的数据基于Range进行Partition的分配,直接写入Shuffle文件中
- Driver 提交ResultTask,Executor读取Shuffle文件中相同的Partition进行合并(相同的key不做值的合并)、排序
- Driver 接收到ResultTask的值后,最后进行不同的Partition数据合并
转自:csdn
作者:raintungli
|
|