Improve KLL Sketch perf for non-grouped queries

[ZacBlanco]

Description

Defers calculations of memory sizes for Kll sketches unless using a grouped query.

Motivation and Context

The current method for calculating memory usage has a hidden cost. Within getEstimatedKllInMemorySize we call getSerializedSizeBytes. The code for the serialized bytes size actually serializes the entire internal state to a byte array first before returning the length. This is expensive and should be avoided.

I am working on a PR to the upstream library to add a less-costly method but until released, I would like to fix this as non-grouped execution doesn't need the memory accounting for every sketch input.

Impact

N/A

Test Plan

N/A

Contributor checklist

• Please make sure your submission complies with our development, formatting, commit message, and attribution guidelines.
• 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

== NO RELEASE NOTE ==

[aaneja]

Can you paste the call stack or method in datasketches-java that ends up serializing the internal state for size calculation ?

[ZacBlanco]

You can find the relevant section here

However, for the KllDoublesSketch, the copying doesn't occur. The call chain eventually gets to here where getNumRetained() is just a few array index lookups. I have a local branch where I'm trying to replace the implementation for doubles with this version of the sketch.

They don't have a sketch implementation for raw longs yet, but I am planning on contributing that to the library so that for the native numeric types we can get better performance creating the sketches. You can see some of the perf numbers I was getting in this comment