Reduction Functions

[gatesn]

We currently limit Vortex to what we term "linear compute" (performing scalar functions over rows or dropping individual rows).

It would be nice to expand Vortex push-down to support "reductions", that is, unordered pair-wise aggregations over rows (not arbitrary group-by, nothing that requires a shuffle).

This would allow us to push-down min, max, sum, roll ups, etc. which can benefit from sigificant speeds up in dictionary, run-end and other encoding schemes.

[gatesn]

I think we could implement this by passing a reduction mask alongside a filter mask, where each set bit signifies the end of a reduction group.

This forces the assumption in the API that reductions can only occur in adjacent rows, and allows us to separate the creation of a reduction mask from the execution.

By setting the end bit, not the start bit, a "group by ()" reduction can use an AllFalse mask. A "group by ID" reduction can use an AllTrue mask.