feat: add array_min scalar function and associated tests

[dharanad]

Which issue does this PR close?

• Closes Support array_min scalar function #16570

Rationale for this change

What changes are included in this PR?

Implemented ArrayMin Scalar function

Are these changes tested?

Yes, Add logic tests as well

Are there any user-facing changes?

[alamb]

Thank you @dharanad -- this looks very nice.

I suspect this function could be substantially sped up with a specialized implementation but we can defer that to when someone is interested in doing so

[alamb]

I recommend putting this in the same file as https://github.com/apache/datafusion/blob/main/datafusion/functions-nested/src/max.rs

perhaps call it https://github.com/apache/datafusion/blob/main/datafusion/functions-nested/src/min_max.rs 🤔

You may be able to share some code, but more importantly I think it will be clearer that min and max are basically the same function

[dharanad]

Sure, should be quick. Let me move refac

[dharanad]

Thank you @dharanad -- this looks very nice.

I suspect this function could be substantially sped up with a specialized implementation but we can defer that to when someone is interested in doing so

"specialized implementation" sounds interesting, can you please ellaborate more i can try working on it on a sperate issue ?

[alamb]

Thank you @dharanad -- this looks very nice.
I suspect this function could be substantially sped up with a specialized implementation but we can defer that to when someone is interested in doing so

"specialized implementation" sounds interesting, can you please ellaborate more i can try working on it on a sperate issue ?

Well the first thing would be to make a benchmark calling array_min for different types of ListArrays (like LIst and List) with different length list elements

The profile with these instructions https://datafusion.apache.org/library-user-guide/profiling.html#profile-the-benchmark

My guess (needs to be verified by profiling):

• you'll find that creating ScalarValue for the min elements takes substantial amounts of time.
• You could probably do better by keeping an array of the min elements in a Vec<u32> and calling the arrow take kernel https://docs.rs/arrow/latest/arrow/compute/kernels/take/fn.take.html

[dharanad]

Thanks for the details @alamb I will try this out and share my findings

[alamb]

Thanks again @dharanad