Non-distributive Aggregate Functions Cs 764 : Advanced Database Project Report

The ability to eeciently compute multiple related group-bys is critical to On-Line Analytical Processing and multidimensional data analysis. The computation of CUBE, a special case of the aggregation problem, has been well studied 1]. However, to the best of our knowledge, previous work focused primarily on computing aggregates over distributive and algebraic functions 3]. In this project, we investigated the possibility of computing "Holistic" aggregate functions eeciently. We started with a straight forward, non-optimized approach and experimented alternatives in various aspects, from retaining more information to engineering for less accurate results. We concluded with evidence that, while hardly any optimization can be made in generating the exact result of holistic functions on the CUBE, an approximation method which we named bucket cube can achieve an accuracy of 85%-99% while incurring only moderately more overhead than computing the CUBE over common distributive functions.