Spark Programming Guide(四)

Shuffle operations

Certain operations within Spark trigger an event known as the shuffle. The shuffle is Spark’s mechanism for re-distributing data so that it’s grouped differently across partitions. This typically involves copying data across executors and machines, making the shuffle a complex and costly operation.

某些操作将会引发spark的“shuffle”操作。“shuffle”是spark重新分配数据的机制，将数据重新分组到不同的分区。这通常涉及到在执行器和机器之间复制数据，使shuffle成为复杂而昂贵的操作。

Background

To understand what happens during the shuffle we can consider the example of the reduceByKey operation. The reduceByKey operation generates a new RDD where all values for a single key are combined into a tuple - the key and the result of executing a reduce function against all values associated with that key. The challenge is that not all values for a single key necessarily reside on the same partition, or even the same machine, but they must be co-located to compute the result

为了能更好的理解“shuffle”过程中发生了什么我们可以以reduceByKey为例来分析
。reduceByKey操作将具备相同键值的元素执行reduce函数聚合到一个tuple中生成一个新的RDD。难题是一个key对应的所有值并不一定在同一个分区里，甚至可能不在同一台机器上，但是它们必须被共同计算。

In Spark, data is generally not distributed across partitions to be in the necessary place for a specific operation. During computations, a single task will operate on a single partition - thus, to organize all the data for a single reduceByKey reduce task to execute, Spark needs to perform an all-to-all operation. It must read from all partitions to find all the values for all keys, and then bring together values across partitions to compute the final result for each key - this is called the shuffle.

在spark中，一些特定的操作需要数据不跨分区分布。在计算期间，一个任务在一个分区上执行，为了所有数据都在单个 reduceByKey 的 reduce 任务上运行，我们需要执行一个 all-to-all 操作。它必须从所有分区读取所有的 key 和 key对应的所有的值，并且跨分区聚集去计算每个 key 的结果 - 这个过程就叫做 shuffle。

Although the set of elements in each partition of newly shuffled data will be deterministic, and so is the ordering of partitions themselves, the ordering of these elements is not. If one desires predictably ordered data following shuffle then it’s possible to use:

尽管每个分区新 shuffle 的数据集将是确定的，分区本身的顺序也是这样，但是这些数据的顺序是不确定的。如果希望 shuffle 后的数据是有序的，可以使用:

mapPartitions to sort each partition using, for example, .sorted
repartitionAndSortWithinPartitions to efficiently sort partitions while simultaneously repartitioning
sortBy to make a globally ordered RDD

mapPartitions 对每个 partition 分区进行排序，例如, .sorted
repartitionAndSortWithinPartitions 在分区的同时对分区进行高效的排序.
sortBy 对 RDD 进行全局的排序

Operations which can cause a shuffle include repartition operations like repartition and coalesce, ‘ByKey operations (except for counting) like groupByKey and reduceByKey, and join operations like cogroup and join.

触发的 shuffle 操作包括 repartition 操作，如 repartition 和 coalesce, ‘ByKey 操作 (除了 counting 之外) 像 groupByKey 和 reduceByKey, 和 join 操作, 像 cogroup 和 join.

Performance Impact

The Shuffle is an expensive operation since it involves disk I/O, data serialization, and network I/O. To organize data for the shuffle, Spark generates sets of tasks - map tasks to organize the data, and a set of reduce tasks to aggregate it. This nomenclature comes from MapReduce and does not directly relate to Spark’s map and reduce operations

Shuffle是一个代价高昂的操作，涉及了磁盘I/O，数据序列化，网络I/O。为了组织shuffle的数据，Spark将生成任务集 - map任务组织数据，reduce任务做数据的聚合。这些术语来自 MapReduce，跟 Spark 的 map 操作和 reduce 操作没有关系。

Internally, results from individual map tasks are kept in memory until they can’t fit. Then, these are sorted based on the target partition and written to a single file. On the reduce side, tasks read the relevant sorted blocks.

在内部，一个 map 任务的所有结果数据会保存在内存，直到内存不能全部存储为止。然后，这些数据将基于目标分区进行排序并写入一个单独的文件中。在 reduce 时，任务将读取相关的已排序的数据块。

Certain shuffle operations can consume significant amounts of heap memory since they employ in-memory data structures to organize records before or after transferring them. Specifically, reduceByKey and aggregateByKey create these structures on the map side, and ‘ByKey operations generate these on the reduce side. When data does not fit in memory Spark will spill these tables to disk, incurring the additional overhead of disk I/O and increased garbage collection.

某些 shuffle 操作会大量消耗堆内存空间，因为 shuffle 操作在数据转换前后，需要在使用内存中的数据结构对数据进行组织。需要特别说明的是，reduceByKey 和 aggregateByKey 在 map 时会创建这些数据结构，’ByKey 操作在 reduce 时创建这些数据结构。当内存满的时候，Spark 会把溢出的数据存到磁盘上，这将导致额外的磁盘 I/O 开销和垃圾回收开销的增加。

Shuffle also generates a large number of intermediate files on disk. As of Spark 1.3, these files are preserved until the corresponding RDDs are no longer used and are garbage collected. This is done so the shuffle files don’t need to be re-created if the lineage is re-computed. Garbage collection may happen only after a long period of time, if the application retains references to these RDDs or if GC does not kick in frequently. This means that long-running Spark jobs may consume a large amount of disk space. The temporary storage directory is specified by the spark.local.dir configuration parameter when configuring the Spark context.

shuffle 操作还会在磁盘上生成大量的中间文件。在 Spark 1.3 中，这些文件将会保留至对应的 RDD 不在使用并被垃圾回收为止。这么做的好处是，如果在 Spark 重新计算 RDD 的血统关系（lineage）时，shuffle 操作产生的这些中间文件不需要重新创建。如果 Spark 应用长期保持对 RDD 的引用，或者垃圾回收不频繁，这将导致垃圾回收的周期比较长。这意味着，长期运行 Spark 任务可能会消耗大量的磁盘空间。临时数据存储路径可以通过 SparkContext 中设置参数 spark.local.dir 进行配置。

Shuffle behavior can be tuned by adjusting a variety of configuration parameters. See the ‘Shuffle Behavior’ section within the Spark Configuration Guide.

shuffle 操作的行为可以通过调节多个参数进行设置。详细的说明请看 Spark 配置指南 中的 “Shuffle 行为” 部分。