ulysses-you commented on a change in pull request #33188:
URL: https://github.com/apache/spark/pull/33188#discussion_r672779057
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/exchange/EnsureRequirements.scala
##########
@@ -250,7 +250,12 @@ object EnsureRequirements extends Rule[SparkPlan] {
def apply(plan: SparkPlan): SparkPlan = plan.transformUp {
// TODO: remove this after we create a physical operator for
`RepartitionByExpression`.
- case operator @ ShuffleExchangeExec(upper: HashPartitioning, child, _) =>
+ // SPARK-35989: AQE will change the partition number so we should retain
the REPARTITION_BY_NUM
+ // shuffle which is specified by user. And also we can not remove
REBALANCE_PARTITIONS_BY_COL,
+ // it is a special shuffle used to rebalance partitions.
+ // So, here we only remove REPARTITION_BY_COL in AQE.
+ case operator @ ShuffleExchangeExec(upper: HashPartitioning, child,
shuffleOrigin)
+ if shuffleOrigin == REPARTITION_BY_COL ||
!conf.adaptiveExecutionEnabled =>
Review comment:
yeah, as we have only skipped applying `CoalesceShufflePartitions` or
other custom shuffle reader at final stage. But for the stages which are in the
process, we do nothing. That's why I think a little bit hack.
One other hack idea is we can remark the shuffle which is before the removed
shuffle and change the `ENSURE_REQUIREMENTS` to `REPARTITION_BY_COL`. Then in
AQE, we can do optimization safely.
IMO, I prefer the idea of `skip removing shuffle with all shuffle origin in
AQE`, it's simple and it can be seen as a behavior change due to AQE is enabled
by default. If user really hit this issue, they can just disable AQE.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
