[GLUTEN-4668][CH] Merge two phase hash-based aggregate into one aggregate in the spark plan when there is no shuffle #4669
Conversation
|
Run Gluten Clickhouse CI |
|
@liujiayi771 @lgbo-ustc please help to review, thanks. |
|
Run Gluten Clickhouse CI |
| } | ||
| } | ||
|
|
||
| override def apply(plan: SparkPlan): SparkPlan = { |
There was a problem hiding this comment.
It is better to use PhysicalPlanSelector.maybe to check if guten is enabled.
There was a problem hiding this comment.
Perhaps you should check whether the SparkPlan has been tagged as TRANSFORM_UNSUPPORTED. There are some rules earlier that will tag the SparkPlan, for example, FallbackOnANSIMode will tag all SparkPlans as TRANSFORM_UNSUPPORTED when ANSI mode is enabled. In this case, we can also avoid merging the aggregations.
|
Run Gluten Clickhouse CI |
| if (isPartialAgg(child, hashAgg)) { | ||
| // convert to complete mode aggregate expressions | ||
| val completeAggregateExpressions = aggregateExpressions.map(_.copy(mode = Complete)) | ||
| HashAggregateExec( |
There was a problem hiding this comment.
Can we use
hashAgg.copy(
aggregateExpressions = completeAggregateExpressions,
child = child.child
)
|
Run Gluten Clickhouse CI |
zzcclp
force-pushed
the
merge_hashagg_if_no_shuffle
branch
from
7aa0f55
to
35c2555
Compare
|
Run Gluten Clickhouse CI |
2 similar comments
|
Run Gluten Clickhouse CI |
|
Run Gluten Clickhouse CI |
zzcclp
force-pushed
the
merge_hashagg_if_no_shuffle
branch
from
46d6a6d
to
17c9586
Compare
|
Run Gluten Clickhouse CI |
|
@liujiayi771 @lgbo-ustc @ulysses-you @PHILO-HE @rui-mo please help to review again, thanks. |
zzcclp
force-pushed
the
merge_hashagg_if_no_shuffle
branch
from
17c9586
to
f5e1689
Compare
|
Run Gluten Clickhouse CI |
| resultExpressions, | ||
| child: HashAggregateExec) | ||
| if !isStreaming && isTransformable(hashAgg) && isTransformable(child) => | ||
| if (isPartialAgg(child, hashAgg)) { |
There was a problem hiding this comment.
why not put this if into previous line ?
| } else { | ||
| objectHashAgg | ||
| } | ||
| case plan: SparkPlan => plan |
There was a problem hiding this comment.
can we also handle SortAggregate ? it is possible that there is no shuffle and sort between two SortAggregate.
| // to handle outputs according to the AggregateMode | ||
| for (attr <- child.output) { | ||
| typeList.add(ConverterUtils.getTypeNode(attr.dataType, attr.nullable)) | ||
| nameList.add(ConverterUtils.genColumnNameWithExprId(attr)) | ||
| nameList.addAll(ConverterUtils.collectStructFieldNames(attr.dataType)) | ||
| } | ||
| (child.output, output) | ||
| } else if (!modes.contains(Partial)) { | ||
| } else if (modes.forall(_ == Partial)) { |
There was a problem hiding this comment.
I'm not sure how CH backend transform aggregate. But the code seems different with before, Partial can appear with PartialMerge with distinct.
zzcclp
force-pushed
the
merge_hashagg_if_no_shuffle
branch
from
f5e1689
to
51e5270
Compare
|
Run Gluten Clickhouse CI |
Updated, please review again, thanks. @ulysses-you @rui-mo |
| agg.resultExpressions, | ||
| agg.child | ||
| ) | ||
| transformer.doValidate().isValid |
There was a problem hiding this comment.
Here we have not pulled out pre/post project, so this validation seems very likely to fail. Do we need this check ? I think it should be fine even we fallback to vanilla Spark after merging aggregates, since there already exists a rule ReplaceHashWithSortAgg to do the similar thing.
There was a problem hiding this comment.
I am not sure that merging sort aggregates also works for the vanilla Spark, do you have any idea for this? the hash based aggregates may be work even fallback.
zzcclp
force-pushed
the
merge_hashagg_if_no_shuffle
branch
from
51e5270
to
580338f
Compare
|
Run Gluten Clickhouse CI |
|
lgtm if test pass, cc @rui-mo if you have other comments |
…gate in the spark plan when there is no shuffle
Examples:
HashAggregate(t1.i, SUM, final)
| => HashAggregate(t1.i, SUM, complete)
HashAggregate(t1.i, SUM, partial)
now this feature only support for CH backend.
Close apache#4668.
Co-authored-by: lgbo <lgbo.ustc@gmail.com>
zzcclp
force-pushed
the
merge_hashagg_if_no_shuffle
branch
from
580338f
to
66bf015
Compare
|
Run Gluten Clickhouse CI |
|
LGTM |
|
===== Performance report for TPCH SF2000 with Velox backend, for reference only ====
|
What changes were proposed in this pull request?
Merge two phase hash-based aggregate into one aggregate in the spark plan when there is no shuffle between them:
Examples:
For example:
TPCH Q18 with bucket tables, before this pr:
there are two
HashAggregateTransformerin one whole stage;after this pr:
there is only one
HashAggregateTransformerin one whole stage, and will reduce the time for the secondHashAggregateTransformer.Now this feature only support for CH backend.
Close #4668.
(Fixes: #4668)
How was this patch tested?
(Please explain how this patch was tested. E.g. unit tests, integration tests, manual tests)
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)