[ 
https://issues.apache.org/jira/browse/CALCITE-5177?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17550215#comment-17550215
 ] 

Ruben Q L commented on CALCITE-5177:
------------------------------------

[~julianhyde] I totally agree with you. 
There are a few places where the new RelNode will get the source hints 
transparently, such as the relevant RelNodes' copy methods (e.g. [here in 
LogicalJoin|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/rel/logical/LogicalJoin.java#L185]),
 or when a rule gets applied (via 
[RelOptRuleCall#transformTo|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/plan/RelOptRuleCall.java#L269]).
However, in other transformations (such as ReflectiveVisitors like 
RelFieldTrimmer, RelStructuredTypeFlattener or RelDecorrelator), the new 
RelNode is created otherwise, e.g. recreated "from scratch" by using a 
RelBuilder, or with LogicalXXX#create. In such scenarios, as of today, the 
hints need to be explicitly transferred from the old node to the new, which is 
a bit fragile and an error-prone situation.
Perhaps we could improve this situation inside {{RelDecorrelator}} by 
centralizing the hints transfer in a single location: inside the {{register}} 
method (which has the "old" and "new" nodes as parameters: if both are hintable 
then transfer the hints from the former to the latter, or something like that), 
instead of each individual {{decorrelateRel}} method.
The same idea might be applicable for {{RelStructuredTypeFlattener}} (using 
{{setNewForOldRel}} method).
Not sure yet how we could simplify RelFielTrimer...

> Query loses hint after decorrelation
> ------------------------------------
>
>                 Key: CALCITE-5177
>                 URL: https://issues.apache.org/jira/browse/CALCITE-5177
>             Project: Calcite
>          Issue Type: Bug
>          Components: core
>            Reporter: Ruben Q L
>            Priority: Blocker
>             Fix For: 1.31.0
>
>
> This seems to be a regression caused by CALCITE-5107, which enlarged the list 
> of hintable RelNodes by making Filter, SetOp, Sort, Window, Values hintable.
> However, it seems that this patch "missed" some Calcite preprocessing 
> elements that can change a plan, and therefore require a manual / ad-hoc 
> treatment of hints to avoid losing them in their processing, e.g. 
> RelDecorrelator.
> In my specific example, let's have this query:
> {code:sql}
> SELECT /*+ MY_HINT_FOR_JOIN */ c.c_custkey, , c.c_name FROM customer c
> WHERE NOT EXISTS (
>   SELECT 1 FROM orders o WHERE o.o_custkey = c.c_custkey AND o.o_orderstatus 
> <> 'abc'
> ) ORDER BY c.c_custkey
> {code}
> Before CALCITE-5107, when the query was parsed, a logical plan was created, 
> my hint was attached to a LogicalProject. The plan contained a 
> LogicalCorrelate (to implement the NOT EXISTS). The plan was then 
> decorrelated with RelDecorrelator, and as a result, we obtained a new plan 
> with a LogicalJoin (instead of correlate), where the hint had been propagated 
> from the projection until the join. Everything is fine.
> After CALCITE-5107, when the query was parsed, now we obtain the same logical 
> plan, but now the hint is attached to the LogicalSort (not to the 
> LogicalProject). When the decorrelator is executed, the plan is transformed 
> (to have LogicalJoin), but {*}the hint is lost{*}, it is not in the new 
> plan's sort (or project, or join, it's nowhere).
> Running the debugger, it seems the problem is inside 
> RelDecorrelator#decorrelateQuery:
> {code:java}
> ...
>   if (!decorrelator.cm.mapCorToCorRel.isEmpty()) {
>     newRootRel = decorrelator.decorrelate(newRootRel); // <-- HINT LOST HERE!
>   }
>   // NOTHING GETS PROPAGATED BECAUSE THE HINT WAS LOST!
>   newRootRel = RelOptUtil.propagateRelHints(newRootRel, true);
>   
>   return newRootRel;
> {code}
> The root cause seems to be that, inside [RelDecorrelator's 
> code|https://github.com/apache/calcite/blob/main/core/src/main/java/org/apache/calcite/sql2rel/RelDecorrelator.java],
>  there are only a few places where hints are copied from the "old" RelNode to 
> the newly created RelNode: 3 calls to RelOptUtil#copyRelHints 
> ([here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelDecorrelator.java#L589],
>  
> [here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelDecorrelator.java#L685],
>  and 
> [here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelDecorrelator.java#L796]),
>  only for projections and aggregates; + RelBuilder#hints [to copy them for 
> joins|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelDecorrelator.java#L1308].
> To sum up, it seems RelDecorrelator only cared to copy hints for the 
> "traditional hintables", so probably something like that is missing for sorts 
> (and other newly hintable RelNodes added by CALCITE-5107).
> Apart from RelDecorrelator, other classes that might suffer from a similar 
> problem are:
>  - 
> [RelFieldTrimmer|https://github.com/apache/calcite/blob/main/core/src/main/java/org/apache/calcite/sql2rel/RelFieldTrimmer.java]:
>  which currently only propagates hints for projections and aggregates 
> ([here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelFieldTrimmer.java#L535],
>  
> [here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelFieldTrimmer.java#L574],
>  and 
> [here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelFieldTrimmer.java#L1115]);
>  and copies hints for joins 
> ([here|https://github.com/apache/calcite/blob/63a15128a332af7a641b26a3224392960882170a/core/src/main/java/org/apache/calcite/sql2rel/RelFieldTrimmer.java#L909]).
>  - 
> [RelStructuredTypeFlattener|https://github.com/apache/calcite/blob/main/core/src/main/java/org/apache/calcite/sql2rel/RelStructuredTypeFlattener.java],
>  which also seems to copy hints only for the "traditional hintables".



--
This message was sent by Atlassian Jira
(v8.20.7#820007)

Reply via email to