[jira] [Commented] (CALCITE-4334) LITERAL_AGG, an aggregate function that returns a constant value

["Benchao Li (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CALCITE-4334?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17724529#comment-17724529 ] 

Benchao Li commented on CALCITE-4334:
-------------------------------------

I understand the intention and benefit of introducing LITERAL_AGG, what I'm thinking is about the limitations it may introduce, some points are below:

- Does the new LITERAL_AGG affect the predicate deducing and expression simplifying, e.g. constant folding?

- Is the new LITERAL_AGG properly handled in Rel2SqlConverter, since this is an internal aggregate function, we may need to prevent it being translated to other dialects.

> LITERAL_AGG, an aggregate function that returns a constant value
> ----------------------------------------------------------------
>
>                 Key: CALCITE-4334
>                 URL: https://issues.apache.org/jira/browse/CALCITE-4334
>             Project: Calcite
>          Issue Type: Bug
>            Reporter: Julian Hyde
>            Assignee: Julian Hyde
>            Priority: Major
>
> It would be useful to have an aggregate function that returns a constant value, regardless of how many rows are in the group. We propose {{LITERAL_AGG}}, an internal aggregate function. As an aggregate function it has no arguments (meaning that it does not read any columns from the input), but a call will have a RexLiteral (constant expression).
> This aggregate function is internal, so there is no SQL syntax. But if I were to write
> {code}
> SELECT deptno, SUM(sal), true
> FROM Emp
> GROUP BY deptno
> {code}
> I should get the following plan:
> {code}
> Aggregate(group={deptno}, aggCalls=SUM($4), LITERAL_AGG(true))
>   Scan(Emp)
> {code}
> Today, the plan would require an extra {{Project}} to add {{RexLiteral(true)}}. Planner rules have to look for a {{Project}} of a literal on top of an {{Aggregate}}; by removing the {{Project}} we simplify planning.
> For example, when rewriting sub-queries (see {{SubQueryRemoveRule}}) we introduce add "true as indicator" to the SELECT clause of sub-queries. It can be used to detect rows generated by an outer join. If it is an aggregate query, we would have to write "min(true) as indicator", which necessitates an extra {{Project}} below the {{Aggregate}} to provide the  "true" value. The LITERAL_AGG aggregate function allows us to avoid the extra {{Project}}.
> Another example came up during CALCITE-4317. We would like to make {{RelBuilder.aggregate(groupKey())}} throw when given an empty group key and no aggregate calls. (Because it would create an {{Aggregate}} that has zero fields, and that is problematic elsewhere in Calcite.) But we would also like a pattern where an aggregate with an empty group key becomes a constant single-row relational expression. So, {{RelBulder.aggregate(groupKey(), aggregateCall(LITERAL_AGG(true)))}} should generate {{VALUES TRUE}}.
> {{LITERAL_AGG}} uses {{interface SqlStaticAggFunction}}, an interface that can optionally be implemented (or wrapped, via {{SqlAggFunction.unwrap(Class)}}), to make the planner and code-generator aware of its properties.
> The implementation adds a {{List<RexNode> rexList}} field to {{AggregateCall}}. This field is empty in all conventional aggregate functions, and has one element in {{LITERAL_AGG}}.



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