[ 
https://issues.apache.org/jira/browse/CALCITE-5953?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758383#comment-17758383
 ] 

Zoltan Haindrich commented on CALCITE-5953:
-------------------------------------------

I would also find it more natural if SUM would follow the mathematical 
definition (so that an empty sum is 0)  I believe when `SUM` was placed next to 
the other aggregates like `MIN` they have made some generalizations.

I was thinking about possible alternatives to fixing this:

* compute it from multiple aggs
{code:sql}
consider: select SUM(CASE WHEN x=1 THEN 2 ELSE 0 END from t;

select CASE WHEN s>0 THEN 2*s ELSE NULLIF(cnt,0) END
from (select count(1) filter (where x=1) as s,count(1) as cnt from t) tt;

or with ANY_VALUE:

select CASE WHEN s>0 THEN 2*s ELSE a END
from (select count(1) filter (where x=1) as s,any_value(0) as a from t) tt;

{code}
** so we are exchanging 1 agg with a complex CASE into 2 simpler aggs + a CASE 
on top of it
** pro of this approach:
*** it could be extended very easily to also cover the case when ELSE is not 0
*** possibly handle more branches for the same variable - at the cost of 1 more 
aggregation
*** it could possibly extended to rewrite multi branch cases - but every branch 
will cost 1 more aggregation
* consider only `SUM0` to optimize
** ProjectAggregateMergeRule could prepare SUM0 for this
** this fixes all above issues
*** but also disables it from the more natural usages

I'm still thinking about alternatives - let me know what you think about the 
above ; or if you have any other ideas!


> AggregateCaseToFilterRule may make inaccurate SUM transformation
> ----------------------------------------------------------------
>
>                 Key: CALCITE-5953
>                 URL: https://issues.apache.org/jira/browse/CALCITE-5953
>             Project: Calcite
>          Issue Type: Bug
>            Reporter: Zoltan Haindrich
>            Assignee: Zoltan Haindrich
>            Priority: Major
>
> consider: {{sum(case when x = 1 then 2 else 0 end) as b}}
> notice that this expression may only be null if there are no rows in the table
> {{AggregateCaseToFilterRule}} rewrites the above expression to {{sum(1) 
> filter (where x=2)}} which broadens when it could be `null` to when there are 
> no matches to the filter
> * *A* is *0* correctly in this case; but I think it will be still *0* in case 
> there are 0 input rows
> * The result for *B* supposed to be *0* but since there are no matches by the 
> filter it becomes *null*
> * *C* is not touched
> {code}
> # Convert CASE to FILTER without matches
> select  sum(case when x = 1 then 1 else 0 end) as a,
>         sum(case when x = 1 then 2 else 0 end) as b,
>         sum(case when x = 1 then 3 else -1 end) as c
> from (values 0, null, 0, 2) as t(x);
> +---+---+----+
> | A | B | C  |
> +---+---+----+
> | 0 | 0 | -4 |
> +---+---+----+
> (1 row)
> !ok
> EnumerableCalc(expr#0..2=[{inputs}], expr#3=[CAST($t0):INTEGER], A=[$t3], 
> B=[$t1], C=[$t2])
>   EnumerableAggregate(group=[{}], A=[COUNT() FILTER $1], B=[SUM($2) FILTER 
> $3], C=[SUM($0)])
>     EnumerableCalc(expr#0=[{inputs}], expr#1=[1], expr#2=[=($t0, $t1)], 
> expr#3=[3], expr#4=[-1], expr#5=[CASE($t2, $t3, $t4)], expr#6=[IS TRUE($t2)], 
> expr#7=[2], $f2=[$t5], $f3=[$t6], $f4=[$t7], $f5=[$t6])
>       EnumerableValues(tuples=[[{ 0 }, { null }, { 0 }, { 2 }]])
> !plan
> {code}



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

Reply via email to