[
https://issues.apache.org/jira/browse/CALCITE-6357?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17835868#comment-17835868
]
Kenneth Knowles commented on CALCITE-6357:
------------------------------------------
[~julianhyde] - are we "holding it wrong"? Can you help us understand how?
I presume you would agree that names of output columns is as much part of data
integrity as the values. So in BeamAggregateProjectMergeRule are we somehow
making Calcite think this is not the output column?
It does seem that our rule is the delta that causes the problem. Calcite
without that rule seems to do the right thing. The rule does one thing: runs
the normal AggregateProjectMergeRule only when the underlying connector does
*not* support pushdown.
In the queries that hit this bug, we know that the connector does *not* support
pushdown, so the original behavior should be maintained.
> Calcite enforces select arguments count to be same as row schema fields which
> causes aliases to be ignored
> ----------------------------------------------------------------------------------------------------------
>
> Key: CALCITE-6357
> URL: https://issues.apache.org/jira/browse/CALCITE-6357
> Project: Calcite
> Issue Type: Bug
> Reporter: Brachi Packter
> Priority: Major
>
> Calcite RelBuilder.ProjectNamed checks if row size in the select is identical
> to schema fields, if no, it creates a project with fields as they appear in
> the select , meaning if they have aliases, they are returning with their
> aliases.
> Here, it checks if they are identical:
> https://github.com/apache/calcite/blob/f14cf4c32b9079984a988bbad40230aa6a59b127/core/src/main/java/org/apache/calcite/tools/RelBuilder.java#L2063
> using RexUtil.isIdentity method:
> ```
> public static boolean isIdentity(List<? extends RexNode> exps,
> RelDataType inputRowType) {
> return inputRowType.getFieldCount() == exps.size()
> && containIdentity(exps, inputRowType, Litmus.IGNORE);
> }
> ```
> This is the problematic part `inputRowType.getFieldCount() == exps.size()`
> If they are identical, and return with their aliases, it is ignored in the
> "rename" method later on
> https://github.com/apache/calcite/blob/f14cf4c32b9079984a988bbad40230aa6a59b127/core/src/main/java/org/apache/calcite/tools/RelBuilder.java#L2125
> and alias is skipped
> https://github.com/apache/calcite/blob/f14cf4c32b9079984a988bbad40230aa6a59b127/core/src/main/java/org/apache/calcite/tools/RelBuilder.java#L2137
> This doesn't impact calcite queries, but in Apache Beam they are doing some
> optimization on top of it,
> https://github.com/apache/beam/blob/master/sdks/java/extensions/sql/src/main/java/org/apache/beam/sdk/extensions/sql/impl/rule/BeamAggregateProjectMergeRule.java
> which causes aliases to be ignored, and data is returning suddenly without
> correct column field.
> I believe the isIdentity check can causes more issues if not fixed, we need
> to understand why is it enforced? isn't it valid to have different size of
> fields in select from what we have in the schema?
> In our case we have a one big row and we run on it different queries, each
> with different fields in the select.
> Beam issue
> https://github.com/apache/beam/issues/30498
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
