[
https://issues.apache.org/jira/browse/FLINK-41516?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18146322#comment-18146322
]
Gustavo de Morais commented on FLINK-41516:
-------------------------------------------
Fixed in apache/flink:master 4721ddb03fbfe183a56fe525af91c8ee6d4095ab
> Row-semantic PTFs keep an unnecessary ChangelogNormalize for upsert sources
> with full deletes
> ---------------------------------------------------------------------------------------------
>
> Key: FLINK-41516
> URL: https://issues.apache.org/jira/browse/FLINK-41516
> Project: Flink
> Issue Type: Bug
> Components: Table SQL / Planner
> Reporter: Gustavo de Morais
> Assignee: Gustavo de Morais
> Priority: Major
> Labels: pull-request-available
>
> Since FLINK-41371, row-semantic PTFs with insert-only output consume upsert
> inputs directly. For an upsert source with full deletes, the planner still
> keeps a {{ChangelogNormalize}} below the PTF, even though it emits the same
> changelog it receives.
> The delete-kind inference for PTF table arguments first requests key-only
> deletes. The {{ChangelogNormalize}} answers with full deletes, which does not
> satisfy the request, but the result is accepted without a check, so the
> fallback to full deletes - which would remove the normalize - never runs.
> {{visitChildren}} performs this check; the PTF branch calls {{visit}}
> directly. Joins use the same pattern but are not affected, because an
> {{Exchange}} always sits between them and the normalize.
> We should only accept the key-only result if it satisfies the requested
> trait, as {{visitChildren}} does.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)