Github user yhuai commented on a diff in the pull request: https://github.com/apache/spark/pull/5208#discussion_r28395812 --- Diff: sql/core/src/main/scala/org/apache/spark/sql/execution/Exchange.scala --- @@ -157,28 +205,61 @@ private[sql] case class AddExchange(sqlContext: SQLContext) extends Rule[SparkPl case Seq(a,b) => a compatibleWith b }.exists(!_) - // Check if the partitioning we want to ensure is the same as the child's output - // partitioning. If so, we do not need to add the Exchange operator. - def addExchangeIfNecessary(partitioning: Partitioning, child: SparkPlan): SparkPlan = - if (child.outputPartitioning != partitioning) Exchange(partitioning, child) else child + // Adds Exchange or Sort operators as required + def addOperatorsIfNecessary( + partitioning: Partitioning, + rowOrdering: Seq[SortOrder], + child: SparkPlan): SparkPlan = { + val needSort = rowOrdering.nonEmpty && child.outputOrdering != rowOrdering + val needsShuffle = child.outputPartitioning != partitioning + val canSortWithShuffle = Exchange.canSortWithShuffle(partitioning, rowOrdering) + + if (needSort && needsShuffle && canSortWithShuffle) { + Exchange(partitioning, rowOrdering, child) + } else { + val withShuffle = if (needsShuffle) { + Exchange(partitioning, Nil, child) + } else { + child + } - if (meetsRequirements && compatible) { + val withSort = if (needSort) { + Sort(rowOrdering, global = false, withShuffle) + } else { + withShuffle + } + + withSort + } + } + + if (meetsRequirements && compatible && !needsAnySort) { operator } else { // At least one child does not satisfies its required data distribution or // at least one child's outputPartitioning is not compatible with another child's // outputPartitioning. In this case, we need to add Exchange operators. - val repartitionedChildren = operator.requiredChildDistribution.zip(operator.children).map { - case (AllTuples, child) => - addExchangeIfNecessary(SinglePartition, child) - case (ClusteredDistribution(clustering), child) => - addExchangeIfNecessary(HashPartitioning(clustering, numPartitions), child) - case (OrderedDistribution(ordering), child) => - addExchangeIfNecessary(RangePartitioning(ordering, numPartitions), child) - case (UnspecifiedDistribution, child) => child - case (dist, _) => sys.error(s"Don't know how to ensure $dist") + val requirements = + (operator.requiredChildDistribution, operator.requiredChildOrdering, operator.children) + + val fixedChildren = requirements.zipped.map { + case (AllTuples, rowOrdering, child) => + addOperatorsIfNecessary(SinglePartition, rowOrdering, child) + case (ClusteredDistribution(clustering), rowOrdering, child) => + addOperatorsIfNecessary(HashPartitioning(clustering, numPartitions), rowOrdering, child) + case (OrderedDistribution(ordering), rowOrdering, child) => + addOperatorsIfNecessary(RangePartitioning(ordering, numPartitions), Nil, child) --- End diff -- Since we have already have all of the needed functions, why not put the `rowOrdering` back (if it is indeed wrong to ignore it)? If we leave it as is, in future maybe a new `SparkPlan` requires both OrderedDistribution and some kind of row ordering (for example, using range partitioner to handle data skew), then the physical plan will be wrong.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org