[
https://issues.apache.org/jira/browse/CALCITE-7058?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17980478#comment-17980478
]
Mihai Budiu commented on CALCITE-7058:
--------------------------------------
I thought that RelDecorrelator was the way to go. I think [~suibianwanwan33]
has spent quite some time trying to improve it.
I am also planning to start working on it more seriously.
Maybe we should discuss this as part of CALCITE-7031?
I think that the general decorrelator cannot be simply implemented as a series
of planner rules, because the algorithm has a precondition that the left-hand
side relation or a correlate has a specific shape (not a multiset). This can be
enforced by having a recursive implementation, as in the RelDecorrelator. But I
don't think this can be checked syntactically, in a normal planner rule. But in
my mind a general decorrelator is based on a recursive implementation which
invokes standard plan rewrite rules.
> Decorrelator may produce different column names
> -----------------------------------------------
>
> Key: CALCITE-7058
> URL: https://issues.apache.org/jira/browse/CALCITE-7058
> Project: Calcite
> Issue Type: Bug
> Components: core
> Affects Versions: 1.40.0
> Reporter: Mihai Budiu
> Assignee: Mihai Budiu
> Priority: Minor
>
> In [CALCITE-7024] I have introduced an assertion (Litmus) check in the
> decorrelator that the output relation type is identical to the input relation
> type. This assertion is too strong, because the decorrelator sometimes
> produces different column names in the result.
> This may be a bug, but in the meantime it can be side-stepped by making the
> check a bit weaker: the types are identical up to column names.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
