-
Notifications
You must be signed in to change notification settings - Fork 5.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add HBO support for CTE materialization #22606
Conversation
public Optional<PlanNode> visitSequence(SequenceNode node, Context context) | ||
{ | ||
node.getCteProducers().forEach(x -> x.accept(this, context)); | ||
return node.getPrimarySource().accept(this, context); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For sequence node, just return the canonicalized primary source
@Override | ||
public Optional<PlanNode> visitCteProducer(CteProducerNode node, Context context) | ||
{ | ||
return node.getSource().accept(this, context); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For cte producer, canonicalize its source
@Override | ||
public Optional<PlanNode> visitCteConsumer(CteConsumerNode node, Context context) | ||
{ | ||
return node.getOriginalSource().accept(this, context); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Canonicalize original source for cte consumer
@Override | ||
public Optional<PlanNode> visitSequence(SequenceNode node, Context context) | ||
{ | ||
node.getCteProducers().forEach(x -> x.accept(this, context)); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Do we need to visit producers?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, it's to trigger canonicalization of plan nodes under the cte producer, which will also need HBO to optimize the subquery.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good, just note that the stats equivalent node of the cteProducer and cteConsumer is the actual source too, I think this stats equivalent node is also passed to the table writers and table scan nodes in the physical optimizer and so tracking should also work but the flow is not entirely clear.
@@ -830,6 +833,25 @@ public Optional<PlanNode> visitAggregation(AggregationNode node, Context context | |||
return Optional.of(canonicalPlan); | |||
} | |||
|
|||
@Override | |||
public Optional<PlanNode> visitSequence(SequenceNode node, Context context) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Add a test case where the sequence node is also canonicalized?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The end to end test in testHistoryBasedStatsCalculatorCTE should suffice to make sure that it works.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I was initially thinking that since the matching happens at the PlanFragment level, the Sequence would have been rewritten anyway (fyi - removing this method does not cause the test testHistoryBasedStatsCalculatorCTE to fail). But leaving it in is probably safe and future-proof
.setSystemProperty(CTE_PARTITIONING_PROVIDER_CATALOG, "hive") | ||
.build(); | ||
// CBO Statistics | ||
assertPlan(cteMaterialization, sql, anyTree(node(ProjectNode.class, anyTree(any())).withOutputRowCount(Double.NaN))); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@feilong-liu will this also work if the cte materialization was off in the first execution and then on in the second?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Also the opposite case, if cte materialization was on for the first run and then off again. I am a little confused and not sure it will work as expected in all these cases.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
No, both cases are not supported. History of CTE materialization runs only apply to CTE materialization runs and non cte materialization runs only apply to non cte materialization runs.
The challenge of support history share between CTE and non CTE runs is that, for non CTE materialization runs, the CTE will be inlined and expanded in query plan, which will have different variable name and plan node ids, and may also go through different optimizations which results in different plans. In order to match the plan hash of CTEs between CTE and non CTE runs, we need to carefully design and change the HBO related plan caononicalization strategy and code, which is non-trival.
So I prefer to first enable HBO for history to be used within CTE matrilizations queries. Otherwise HBO will be barely available (as the CTE related node is not canonicalized and will return empty in hash, which will lead to all other operators which has these CTE nodes as child also return empty in hash)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sg, I have fixed the planNodeIds and now they are canonicalized as expected in an earlier commit.
For the different variable names, I believe we can still use the underlying plan statistics available in cteConsumer/producer (original node HBO stats). These statistics will be accessible if the previous query was run in non-CTE materialized mode. However linking with the execution stats is unclear
Suggest minor change in release note entry to include the PR number:
|
Description
The three nodes, Sequence node, cte producer node, and cte consumer node are not handled in Canonicalization for HBO. It will make HBO not working for the query. This PR adds plan canonicalization for it.
Motivation and Context
To have HBO work for CTE materialized query
Impact
HBO work for CTE materialized query
Test Plan
Unit test
Contributor checklist
Release Notes
Please follow release notes guidelines and fill in the release notes below.