[Design] Consider More General Aggregation Decomposition

[Selvomega]

Problem and why now

Currently ASAPPlanner only break down aggregations in an ad-hoc manner, say, breaking avg into count and sum.
We argue for a more comprehensive design of such decomposition from non-mergeable aggregation to mergeable ones.

Such decomposition will help

• reduce computation redundancy by sharing "basic building blocks"
• Maintain incremental computation (since the decomposed states are mergeable)

Intended users and MVP

Intended users can be all 4 down-stream use cases of ASAPPlanner.

For MVP, we can design an aggregation benchmark containing such decomposable aggregations to show the before&after of this design.

Inputs, outputs, and end-to-end outcome

I am visioning this functionality as a small unit in ASAPPlanner with a single aggregation as input, decomposed aggregations as output, say:

• avg in, sum + count out
• var in, sum + square_sum out
• linear_regression in, count + sum_x + sum_y + sum_xy + sum_x_square + sum_y_square out

Simplest viable approach

Adding an extra step to process all aggregations

New concepts or interfaces requiring decisions

I suggest a meeting to crowd-source all decompositions we want.

Alternatives and quality attributes

Acceptance behavior

Human decisions required