[ 
https://issues.apache.org/jira/browse/CALCITE-5193?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17866241#comment-17866241
 ] 

Viggo Chen edited comment on CALCITE-5193 at 7/16/24 6:09 AM:
--------------------------------------------------------------

[~Chunwei Lei] It is intuitive to pushdown `COALESCE(a.pt, b.pt)='20220601'` as 
`pt='20220601'` filter to each input. 
But how do we abstractly judge whether a filter condition can be pushed down? 
I have a preliminary idea, but I'm not sure if it's correct. For left input, 
replace all field from right input as null to create a new filter to pushdown. 
And the same applies to the right side.


was (Author: JIRAUSER300637):
[~Chunwei Lei] It is intuitive to pushdown `COALESCE(a.pt, b.pt)='20220601'` as 
`pt='20220601'` filter to each input. 
But how do we abstractly judge whether a filter condition can be pushed down? 
I have a preliminary idea, but I'm not sure if it's correct. For left input, 
replace all field from right input as null to create a new filter to pushdown. 
And the same applies to the right side.
 
 

> Push filter whose conditions include join keys and are composed by OR into 
> inputs of full join
> ----------------------------------------------------------------------------------------------
>
>                 Key: CALCITE-5193
>                 URL: https://issues.apache.org/jira/browse/CALCITE-5193
>             Project: Calcite
>          Issue Type: Improvement
>            Reporter: Chunwei Lei
>            Priority: Major
>
> For example,
> {code:sql}
> select * from a full join b on a.id=b.id where a.id=1 or b.id=2;
> {code}
> can be transformed to 
> {code:sql}
> select * from 
> (select * from a where id=1 or id=2) a 
> full join 
> (select * from b where id=1 or id=2) b
> on a.id=b.id;
> {code}
> If {{a}} and {{b}} are both partitioned tables and id is the partition key, 
> we can do partition pruning with this transformation, which is a big 
> improvement.
> This improvement is inspired by the query 
> {code:java}
> select * from a full join b on a.id=b.id and a.pt=b.pt where COALESCE(a.pt, 
> b.pt)='20220601';
> {code}
> which costs a lot due to it scans all partitions in table {{a}} and {{b}}.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to