Add optimizer to convert min_by/max_by to row number function

[feilong-liu]

Description

This optimization converts queries like

select id, max(ds), max_by(feature1, ds), max_by(feature2, ds) from table group by id

to

select id, ds, feature1, feature2 from (select id, ds, feature1, feature2, row_number() over (partition by id order by ds desc) row_num) where row_num = 1

Here feature1, feature2 are maps. This rewrite can avoid the expensive cost of aggregations on feature1 and feature2. This is commonly used in getting latest features in machine learning workload.

Motivation and Context

Query optimization to reduce cost.

Impact

Query optimization to reduce cost.

Test Plan

Unit tests

Contributor checklist

• Please make sure your submission complies with our contributing guide, in particular code style and commit standards.
• PR description addresses the issue accurately and concisely. If the change is non-trivial, a GitHub Issue is referenced.
• Documented new properties (with its default value), SQL syntax, functions, or other functionality.
• If release notes are required, they follow the release notes guidelines.
• Adequate tests were added if applicable.
• CI passed.

Release Notes

Please follow release notes guidelines and fill in the release notes below.

== RELEASE NOTES ==

General Changes
* Add a new optimization `MinMaxByToWindowFunction` to rewrite min_by/max_by aggregations with row_number window function

[jaystarshot]

Maybe I am missing something but In
select id, max(ds), max_by(feature1, ds), max_by(feature2, ds) from table group by id
feature1/feature2 are not aggregated right? (they just fetch the feature from the max ds row?) so i don't understand

Here feature1, feature2 are maps. This rewrite can avoid the expensive cost of aggregations on feature1 and feature2.

[feilong-liu]

Maybe I am missing something but In select id, max(ds), max_by(feature1, ds), max_by(feature2, ds) from table group by id feature1/feature2 are not aggregated right? (they just fetch the feature from the max ds row?) so i don't understand

Here feature1, feature2 are maps. This rewrite can avoid the expensive cost of aggregations on feature1 and feature2.

Correct, here is the definition of max_by(x, y) -> [same as x]() https://prestodb.io/docs/current/functions/aggregate.html#max_by-x-y-same-as-x Returns the value of x associated with the maximum value of y over all input values
This is commonly used in feature selection, i.e. to select the latest feature in a table for a user.

[feilong-liu]

Maybe I am missing something but In select id, max(ds), max_by(feature1, ds), max_by(feature2, ds) from table group by id feature1/feature2 are not aggregated right? (they just fetch the feature from the max ds row?) so i don't understand

Here feature1, feature2 are maps. This rewrite can avoid the expensive cost of aggregations on feature1 and feature2.

And the expensive cost of aggregations I mean the process of feature maps in the accumulator (although it's not actually doing aggregation on it by semantics, it can still be expensive, especially for large maps)

[jaystarshot]

Can you add a small comment explaining the plan changes?

[feilong-liu]

Sure, will add in a separate PR.

[feilong-liu]

#25249 @jaystarshot

[jaystarshot]

Can this be on by default?

[jaystarshot]

I guess row number adds sorting so might not be always efficient but if your performance numbers show other wise then we can make it on by default?

[feilong-liu]

I want to be conservative for now. Will consider to set it to be true after getting more stats for this optimizer