[
https://issues.apache.org/jira/browse/HIVE-11152?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14609956#comment-14609956
]
Jesus Camacho Rodriguez commented on HIVE-11152:
------------------------------------------------
[~jpullokkaran], I'm not sure that's the way to go.
The problem that I see is that if we do it over the Calcite plan (on or even
after PlanConverter), and we swap the inputs of the joins accordingly, we would
need to add Project operators on top of the Join operators to keep the plan
correct, thus defeating the main purpose of this change: that the multijoin
optimization kicks in. Otherwise we need to run a whole set of rules again, but
we may be interfering with what the PlanConverter already did.
In turn, I do not think it is possible to "just" swap inputs in Calcite, which
relies on the positions of the fields; this may cause another set of problems
as the upper part of the plan relies on this positions.
Thus, I think it makes sense to add it in ASTConverter. AST relies fully on
column names and not positions, so if the inputs are swapped, nothing will
happen in the upper part of the plan.
--
I checked the fails, and they are either benign or even some improvements (e.g.
correlationoptimizer6.q). I'm uploading a new patch adding the comment to plan
modifier you requested, as well as the new q files.
> Swapping join inputs in ASTConverter
> ------------------------------------
>
> Key: HIVE-11152
> URL: https://issues.apache.org/jira/browse/HIVE-11152
> Project: Hive
> Issue Type: Bug
> Components: CBO
> Reporter: Jesus Camacho Rodriguez
> Assignee: Jesus Camacho Rodriguez
> Attachments: HIVE-11152.patch
>
>
> We want that multijoin optimization in SemanticAnalyzer always kicks in when
> CBO is enabled (if possible). For that, we may need to swap the join inputs
> when we return from CBO through the Hive AST.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
