BypassMergeSortShuffleWriter
就如其名,旁支的sort-baesd Shuffle, 他是采用Hash-style实现的Sort based Shuffle。在map阶段records会按分区写入不同的文件, 一个分区一个文件。然后链接这些分区文件形成一个output文件,并生成其index。reducer通过IndexShuffleBlockResolver
查找消费输出文件的不同分区。
在 BypassMergeSortShuffleWriter
中records是不会缓存在内存中,所有的records最终都会被flush到磁盘。
在写入时,BypassMergeSortShuffleWriter
会同时为所有的分区打开单独的序列化器和文件流,所以当reduce分区数量特别大的时候性能会非常低下。
ShuffleWriter 的调用是在ShuffleMapTask的runTask中进行调用,每个mapTask 都会调用一次runTask。
BypassMergeSortShuffleWriter 源码解析
首先,我们来回顾下ShuffleWriter的过程。Shuffle发生与宽依赖的stage间,由于stage内的计算采用pipeline的方式。shuffle发生的上一个stage为map节点,下游的stage为reduce阶段。而shuffle写的过程就发生在map阶段,shuffleWriter的调用主要是在ShuffleMapStage中,每个ShuffleMapStage包含多个ShuffleMapTask, mapTask个数和分区数相关。
这样每个ShuffleMapTask都会在其runTask调用下Writer接口,其并非直接调用到具体的执行类。而是在划分宽依赖时想ShuffleManage注册shuffle时,返回的ShuffleHandler决定的。
在ShuffleMapTask调用Writer时,是先调用了ShuffleWriteProcessor
,主要控制了ShuffleWriter的生命周期。下面我们看下ShuffleWriteProcessor
中的Write的实现:
// ShuffleWriteProcessor
def write(
rdd: RDD[_],
dep: ShuffleDependency[_, _, _],
mapId: Long,
context: TaskContext,
partition: Partition): MapStatus = {
var writer: ShuffleWriter[Any, Any] = null
try {
// [1] 通过SparkEnv获取ShuffleManager, 并通过dep的shuffleHandle, 获取对应的shuffleWriter的具体实现。
val manager = SparkEnv.get.shuffleManager
writer = manager.getWriter[Any, Any](
dep.shuffleHandle,
mapId,
context,
createMetricsReporter(context))
// [2] 调用shuffleWriter的write方法, 并将当前rdd的迭代器传入
writer.write(
rdd.iterator(partition, context).asInstanceOf[Iterator[_ <: Product2[Any, Any]]])
// [3] shuffleWriter结束后,返回mapStatus,或清空数据
val mapStatus = writer.stop(success = true)
// [4] 如果shuffleWriter执行成功,初始化push-based shuffle, 后面再细讲
if (mapStatus.isDefined) {
// Initiate shuffle push process if push based shuffle is enabled
// The map task only takes care of converting the shuffle data file into multiple
// block push requests. It delegates pushing the blocks to a different thread-pool -
// ShuffleBlockPusher.BLOCK_PUSHER_POOL.
if (dep.shuffleMergeEnabled && dep.getMergerLocs.nonEmpty && !dep.shuffleMergeFinalized) {
manager.shuffleBlockResolver match {
case resolver: IndexShuffleBlockResolver =>
val dataFile = resolver.getDataFile(dep.shuffleId, mapId)
new ShuffleBlockPusher(SparkEnv.get.conf)
.initiateBlockPush(dataFile, writer.getPartitionLengths(), dep, partition.index)
case _ =>
}
}
}
mapStatus.get
}
...
}
ShuffleWriteProcessor
中主要做了三件事:
- [1] 通过SparkEnv获取ShuffleManager, 并通过dep的shuffleHandle, 获取对应的shuffleWriter的具体实现。
- [2] 调用shuffleWriter的write方法, 并将当前rdd的迭代器传入
- [3] shuffleWriter结束后,返回mapStatus,或清空数据
可见每一个ShuffleMapTask执行结束后,就会返回一个mapStatus。Task 结果被封装成 CompletionEvent
发送到Driver DAG Scheduler 。判断Task的类型是ShuffleMapTask会DagScheduler 会向 MapOutputTracker 注册 MapOutput status 信息。
那么map中的数据是如何通过BypassMergeSortShuffleWriter写入的?
代码语言:javascript复制// BypassMergeSortShuffleWriter
@Override
public void write(Iterator<Product2<K, V>> records) throws IOException {
assert (partitionWriters == null);
// [1] 创建处理mapTask所有分区数据commit提交writer
ShuffleMapOutputWriter mapOutputWriter = shuffleExecutorComponents
.createMapOutputWriter(shuffleId, mapId, numPartitions);
try {
// 如果没有数据,直接提交所有分区的commit, 并返回分区长度,获取mapStatus
if (!records.hasNext()) {
partitionLengths = mapOutputWriter.commitAllPartitions(
ShuffleChecksumHelper.EMPTY_CHECKSUM_VALUE).getPartitionLengths();
mapStatus = MapStatus$.MODULE$.apply(
blockManager.shuffleServerId(), partitionLengths, mapId);
return;
}
final SerializerInstance serInstance = serializer.newInstance();
final long openStartTime = System.nanoTime();
// [2] 为每个分区创建一个DiskBlockObjectWriter写入流和FileSegment文件段
partitionWriters = new DiskBlockObjectWriter[numPartitions];
partitionWriterSegments = new FileSegment[numPartitions];
for (int i = 0; i < numPartitions; i ) {
// [2.1] 每个分区创建个临时file和blockid, 并生成维护一个写入流
final Tuple2<TempShuffleBlockId, File> tempShuffleBlockIdPlusFile =
blockManager.diskBlockManager().createTempShuffleBlock();
final File file = tempShuffleBlockIdPlusFile._2();
final BlockId blockId = tempShuffleBlockIdPlusFile._1();
DiskBlockObjectWriter writer =
blockManager.getDiskWriter(blockId, file, serInstance, fileBufferSize, writeMetrics);
if (partitionChecksums.length > 0) {
writer.setChecksum(partitionChecksums[i]);
}
partitionWriters[i] = writer;
}
// Creating the file to write to and creating a disk writer both involve interacting with
// the disk, and can take a long time in aggregate when we open many files, so should be
// included in the shuffle write time.
writeMetrics.incWriteTime(System.nanoTime() - openStartTime);
// [3] 依次将records写入到对应分区的写入流中, 并提交
while (records.hasNext()) {
final Product2<K, V> record = records.next();
final K key = record._1();
partitionWriters[partitioner.getPartition(key)].write(key, record._2());
}
// [3.1]依次对每个分区提交和flush写入流
for (int i = 0; i < numPartitions; i ) {
try (DiskBlockObjectWriter writer = partitionWriters[i]) {
partitionWriterSegments[i] = writer.commitAndGet();
}
}
// [4] 遍历所有分区的FileSegement, 并将其链接为一个文件,同时会调用writeMetadataFileAndCommit,为其生成索引文件
partitionLengths = writePartitionedData(mapOutputWriter);
mapStatus = MapStatus$.MODULE$.apply(
blockManager.shuffleServerId(), partitionLengths, mapId);
} catch (Exception e) {
try {
mapOutputWriter.abort(e);
} catch (Exception e2) {
logger.error("Failed to abort the writer after failing to write map output.", e2);
e.addSuppressed(e2);
}
throw e;
}
}
综上,Bypass的writer步骤有四步:
- [1] 创建处理mapTask所有分区数据commit提交writer
- [2] 为每个分区创建一个DiskBlockObjectWriter写入流和FileSegment文件段
- [2.1] 每个分区创建个临时file和blockid, 并生成维护一个DiskBlockObjectWriter写入流
- [3] 依次将records写入到对应分区的写入流中, 并提交
- [3.1]依次对每个分区提交和flush写入流
- [4] 遍历所有分区的FileSegement, 并将其链接为一个文件,同时会调用writeMetadataFileAndCommit,为其生成索引文件
所以说, Bypass在进行写入时会为每个MapTask都会生成reduce分区个FileSegement, 写入时会并发的为所有的分区都创建临时文件和维护一个io的写入流, 最终在链接为一个文件。所以如果分区数特别多的情况下,是会维护很多io流,所以Bypass限制了分区的阈值。另外通过源码发现Bypass在实现过程中并没有使用buffer, 而是直接将数据写入到流中,这也就是为什么Bypass不能处理mapSide的预聚合的算子。
那么BypassMergeSortShuffleWriter 属于sort-based Shuffle 到底有没有排序呢?
接下来,我们再看下Bypass是如何将分区的FileSegement, 并将其链接为一个文件, 我们就需要详细看下writePartitionedData是如何实现的。
代码语言:javascript复制private long[] writePartitionedData(ShuffleMapOutputWriter mapOutputWriter) throws IOException {
// Track location of the partition starts in the output file
if (partitionWriters != null) {
final long writeStartTime = System.nanoTime();
try {
for (int i = 0; i < numPartitions; i ) {
// [1] 获取每个分区的 fileSegement 临时文件,和writer写出流
final File file = partitionWriterSegments[i].file();
ShufflePartitionWriter writer = mapOutputWriter.getPartitionWriter(i);
if (file.exists()) {
if (transferToEnabled) {
// Using WritableByteChannelWrapper to make resource closing consistent between
// this implementation and UnsafeShuffleWriter.
Optional<WritableByteChannelWrapper> maybeOutputChannel = writer.openChannelWrapper();
if (maybeOutputChannel.isPresent()) {
writePartitionedDataWithChannel(file, maybeOutputChannel.get());
} else {
writePartitionedDataWithStream(file, writer);
}
} else {
// [2] 将fileSegement合并为一个文件
writePartitionedDataWithStream(file, writer);
}
if (!file.delete()) {
logger.error("Unable to delete file for partition {}", i);
}
}
}
} finally {
writeMetrics.incWriteTime(System.nanoTime() - writeStartTime);
}
partitionWriters = null;
}
// [3] 提交所有的分区,传入每个分区数据的长度, 调用 writeMetadataFileAndCommit生成索引文件,记录每个分区的偏移量
return mapOutputWriter.commitAllPartitions(getChecksumValues(partitionChecksums))
.getPartitionLengths();
}
writePartitionedData是如何实现,有三个步骤:
- [1] 获取每个分区的 fileSegement 临时文件,和writer写出流
- [2] 将fileSegement合并为一个文件
- [3] 提交所有的分区,传入每个分区数据的长度, 调用 writeMetadataFileAndCommit生成索引文件,记录每个分区的偏移量
bypass.png
总结, BypassMergeSortShuffleWriter 的实现是hash-style的方式,其中没有sort, 没有buffer,每一个mapTask都会生成分区数量个FileSegment, 最后再合并为一个File, 最终根据分区的长度为其生成索引文件。所以BypassMergeSortShuffleWriter在分区数量比较小的情况下,性能是比较佳的。其最终每个task会生成2个文件, 所以最终的生成文件数也是2 * M个文件。
今天就先到这里,通过上面的介绍,我们也留下些面试题:
- BypassMergeSortShuffleWriter和HashShuffle有什么区别?
- 为什么不保留HashShuffleManage, 而是将其作为SortShuffleManager中的一个Writer实现?