[
https://issues.apache.org/jira/browse/CALCITE-4037?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17169879#comment-17169879
]
Julian Hyde commented on CALCITE-4037:
--------------------------------------
I am skeptical about this bug. Preserving the field names of the Aggregate
seems to be the correct behavior. (As a policy, rules try to preserve field
names, unless doing so would make the result more complex or less efficient.)
SQL-to-Rel seems to have produced some funky field names in this example, but
even so, the rule isn’t the place to fix that.
> AggregateProjectMergeRule doesn't always respect column aliases
> ---------------------------------------------------------------
>
> Key: CALCITE-4037
> URL: https://issues.apache.org/jira/browse/CALCITE-4037
> Project: Calcite
> Issue Type: Bug
> Components: core
> Reporter: Sylvain Crozon
> Priority: Major
> Labels: pull-request-available
> Attachments: fix-AggregateProjectMergeRule.patch
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> When columns aggregated over are given an alias, but not aggregated values,
> the alias is lost:
> {code:java}
> select deptno as x, sum(sal)
> from emp
> group by deptno
> {code}
> has the following plan
> {code:java}
> LogicalAggregate(group=[{0}], EXPR$1=[SUM($1)])
> LogicalProject(X=[$7], SAL=[$5])
> LogicalTableScan(table=[[CATALOG, SALES, EMP]])
> {code}
> which becomes
> {code:java}
> LogicalAggregate(group=[{7}], EXPR$1=[SUM($5)])
> LogicalTableScan(table=[[CATALOG, SALES, EMP]])
> {code}
>
> after the AggregateProjectMergeRule
>
> I attempted a fix by comparing the row type's field names of the project node
> with its input, and skip merging with the agg node if they don't match. That
> works for the use case above, however that breaks quite a few unit tests.
> Some of them, I believe, should be updated like
> testAggregateMerge1, the alias in the SQL queries are lost in the final plan.
> But for other use cases, mostly when there's a join, the simple column name
> comparison is not enough.
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
