Spark 中的 Rebalance 操作以及与Repartition操作的区别

Posted 2022-07-01 鸿乃江边鸟

背景

本文基本spark 3.2.1
在Partitioning Hints Types中有提到Rebalance操作以及Repartition操作，而且他们都可以做数据的重分区，他们之间有什么区别呢？

SELECT /*+ REPARTITION(3) */ * FROM t;

SELECT /*+ REPARTITION(c) */ * FROM t;

SELECT /*+ REBALANCE */ * FROM t;

SELECT /*+ REBALANCE(c) */ * FROM t;

分析

• Rebalance
  参考对应的SPARK-35725,其目的是为了在AQE阶段,根据spark.sql.adaptive.advisoryPartitionSizeInBytes进行分区的重新分区，防止数据倾斜。再加上SPARK-35786,就可以根据hint进行重分区。
  具体看看怎么实现的,OptimizeSkewInRebalancePartitions代码如下：

    override val supportedShuffleOrigins: Seq[ShuffleOrigin] =
     Seq(REBALANCE_PARTITIONS_BY_NONE, REBALANCE_PARTITIONS_BY_COL)
   ...
    override def apply(plan: SparkPlan): SparkPlan = 
   if (!conf.getConf(SQLConf.ADAPTIVE_OPTIMIZE_SKEWS_IN_REBALANCE_PARTITIONS_ENABLED)) 
     return plan
   

   plan match 
     case stage: ShuffleQueryStageExec if isSupported(stage.shuffle) =>
       tryOptimizeSkewedPartitions(stage)
     case _ => plan
   
 

只有开启了 spark.sql.adaptive.optimizeSkewsInRebalancePartitions.enabled了的情况下，才可以进行分区的expand，而且还得shuffle的来源还得是REBALANCE_PARTITIONS_BY_NONE, REBALANCE_PARTITIONS_BY_COL的情况下才能适用该规则.
tryOptimizeSkewedPartitions 的具体实现可以看代码,该代码的注释很清楚：

* We use ADVISORY_PARTITION_SIZE_IN_BYTES size to decide if a partition should be optimized.
* Let's say we have 3 maps with 3 shuffle partitions, and assuming r1 has data skew issue.
* the map side looks like:
*   m0:[b0, b1, b2], m1:[b0, b1, b2], m2:[b0, b1, b2]
* and the reduce side looks like:
*                            (without this rule) r1[m0-b1, m1-b1, m2-b1]
*                              /                                     \\
*   r0:[m0-b0, m1-b0, m2-b0], r1-0:[m0-b1], r1-1:[m1-b1], r1-2:[m2-b1], r2[m0-b2, m1-b2, m2-b2]
*
* Note that, this rule is only applied with the SparkPlan whose top-level node is
* ShuffleQueryStageExec.

我们分析一下REBALANCE_PARTITIONS_BY_NONE, REBALANCE_PARTITIONS_BY_COL来源:
这是在ResolveHints规则中进行转换的:

   private def createRebalance(hint: UnresolvedHint): LogicalPlan = 
    hint.parameters match 
      case partitionExprs @ Seq(_*) =>
        val invalidParams = partitionExprs.filter(!_.isInstanceOf[UnresolvedAttribute])
        if (invalidParams.nonEmpty) 
          val hintName = hint.name.toUpperCase(Locale.ROOT)
          throw QueryCompilationErrors.invalidHintParameterError(hintName, invalidParams)
        
        RebalancePartitions(partitionExprs.map(_.asInstanceOf[Expression]), hint.child)
    
  

  def apply(plan: LogicalPlan): LogicalPlan = plan.resolveOperatorsWithPruning(
    _.containsPattern(UNRESOLVED_HINT), ruleId) 
    case hint @ UnresolvedHint(hintName, _, _) => hintName.toUpperCase(Locale.ROOT) match 
        case "REPARTITION" =>
          createRepartition(shuffle = true, hint)
        case "COALESCE" =>
          createRepartition(shuffle = false, hint)
        case "REPARTITION_BY_RANGE" =>
          createRepartitionByRange(hint)
        case "REBALANCE" if conf.adaptiveExecutionEnabled =>
          createRebalance(hint)
        case _ => hint
      
   

可见只有在AQE开启的情况下 该Rebalance的hint才生效，生成对应的RebalancePartitions逻辑计划，而该逻辑计划会在BasicOperators规则中，转换成ShuffleEchangeExec物理计划：

      case r: logical.RebalancePartitions =>
      val shuffleOrigin = if (r.partitionExpressions.isEmpty) 
        REBALANCE_PARTITIONS_BY_NONE
       else 
        REBALANCE_PARTITIONS_BY_COL
      
      exchange.ShuffleExchangeExec(r.partitioning, planLater(r.child), shuffleOrigin) :: Nil

因为只有shuffle操作的时候，AQE阶段才会应用到OptimizeSkewInRebalancePartitions规则,这样才会在shuffle read阶段根据shuffle write阶段的数据进行优化。
注意:
其中 OptimizeShuffleWithLocalRead 不适用 shuffleOrigin为REBALANCE_PARTITIONS_BY_COL的，要不然在动态分区存在小文件的问题，具体见该处讨论

• Repartition
  相对于Rebalance，该hint只是根据指定的固定的分区数据或者列进行分区，这个时候每个分区的大小并不能控制，只能说是平均分配或者说是按照列进行hash分区（这种情况存在文件大小不一的情况）
  具体的分析，可以参考Rebalance的分析。
  注意一点的是在SPARK-35650之后，Repartition操作也是在AQE阶段进行优化，而在SPARK-35725 之后，如果是单纯的REPARTITION hint 也是可以达到Rebalace hint的效果，因为在此处把shuffleOrigin从REPARTITION_BY_NONE改成了REBALANCE_PARTITIONS_BY_NONE了,所以也能使用于OptimizeSkewInRebalancePartitions规则。

结论

一般在reparition用到的地方都可以Rebalance来替换，而且Rebalance有更好的文件大小的控制能力，更多的信息可以查看对应的 spark-jira