[
https://issues.apache.org/jira/browse/FLINK-33996?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17804557#comment-17804557
]
Feng Jin commented on FLINK-33996:
----------------------------------
[~libenchao] [~qingyue] If it can be reused during the codegen phase, then it
is indeed a better approach.
In addition, what I would like to ask is whether the current optimization rule
will definitely make the same project rewrite into the same cal node, at least
in my actual tests, this is the behavior. otherwise it may not be able to reuse
in the codegen phase.
> Support disabling project rewrite when multiple exprs in the project
> reference the same sub project field.
> ----------------------------------------------------------------------------------------------------------
>
> Key: FLINK-33996
> URL: https://issues.apache.org/jira/browse/FLINK-33996
> Project: Flink
> Issue Type: Improvement
> Components: Table SQL / Runtime
> Affects Versions: 1.18.0
> Reporter: Feng Jin
> Priority: Major
> Labels: pull-request-available
>
> When multiple top projects reference the same bottom project, project rewrite
> rules may result in complex projects being calculated multiple times.
> Take the following SQL as an example:
> {code:sql}
> create table test_source(a varchar) with ('connector'='datagen');
> explan plan for select a || 'a' as a, a || 'b' as b FROM (select
> REGEXP_REPLACE(a, 'aaa', 'bbb') as a FROM test_source);
> {code}
> The final SQL plan is as follows:
> {code:sql}
> == Abstract Syntax Tree ==
> LogicalProject(a=[||($0, _UTF-16LE'a')], b=[||($0, _UTF-16LE'b')])
> +- LogicalProject(a=[REGEXP_REPLACE($0, _UTF-16LE'aaa', _UTF-16LE'bbb')])
> +- LogicalTableScan(table=[[default_catalog, default_database,
> test_source]])
> == Optimized Physical Plan ==
> Calc(select=[||(REGEXP_REPLACE(a, _UTF-16LE'aaa', _UTF-16LE'bbb'),
> _UTF-16LE'a') AS a, ||(REGEXP_REPLACE(a, _UTF-16LE'aaa', _UTF-16LE'bbb'),
> _UTF-16LE'b') AS b])
> +- TableSourceScan(table=[[default_catalog, default_database, test_source]],
> fields=[a])
> == Optimized Execution Plan ==
> Calc(select=[||(REGEXP_REPLACE(a, 'aaa', 'bbb'), 'a') AS a,
> ||(REGEXP_REPLACE(a, 'aaa', 'bbb'), 'b') AS b])
> +- TableSourceScan(table=[[default_catalog, default_database, test_source]],
> fields=[a])
> {code}
> It can be observed that after project write, regex_place is calculated twice.
> Generally speaking, regular expression matching is a time-consuming operation
> and we usually do not want it to be calculated multiple times. Therefore, for
> this scenario, we can support disabling project rewrite.
> After disabling some rules, the final plan we obtained is as follows:
> {code:sql}
> == Abstract Syntax Tree ==
> LogicalProject(a=[||($0, _UTF-16LE'a')], b=[||($0, _UTF-16LE'b')])
> +- LogicalProject(a=[REGEXP_REPLACE($0, _UTF-16LE'aaa', _UTF-16LE'bbb')])
> +- LogicalTableScan(table=[[default_catalog, default_database,
> test_source]])
> == Optimized Physical Plan ==
> Calc(select=[||(a, _UTF-16LE'a') AS a, ||(a, _UTF-16LE'b') AS b])
> +- Calc(select=[REGEXP_REPLACE(a, _UTF-16LE'aaa', _UTF-16LE'bbb') AS a])
> +- TableSourceScan(table=[[default_catalog, default_database,
> test_source]], fields=[a])
> == Optimized Execution Plan ==
> Calc(select=[||(a, 'a') AS a, ||(a, 'b') AS b])
> +- Calc(select=[REGEXP_REPLACE(a, 'aaa', 'bbb') AS a])
> +- TableSourceScan(table=[[default_catalog, default_database,
> test_source]], fields=[a])
> {code}
> After testing, we probably need to modify these few rules:
> org.apache.flink.table.planner.plan.rules.logical.FlinkProjectMergeRule
> org.apache.flink.table.planner.plan.rules.logical.FlinkCalcMergeRule
> org.apache.flink.table.planner.plan.rules.logical.FlinkProjectCalcMergeRule
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
