Foresight is golden: use of prediction in DB systems

 

Dr. Sean Wang

Department of Computer Science

University of Vermont

 

Date: Monday September 13, 2004

Time: 12:20 p.m. - 1:10 p.m.

Location: 367 Votey

 

 

Abstract

 

In many situations, prediction is possible in terms of either user or system behavior. The use of prediction is not new to computer systems. Examples include prefetching in various scenarios, such as in processors, operating systems, and database buffer management. In this talk, we argue that the use of prediction can be at a higher abstraction level as well. As an example, we show a prediction-based method for maintaining a multi-dimensional histogram for selectivity estimation. Almost all histogram methods found in the literature are "reactive" in nature, that is, they are trying to do the best given the past data and/or queries. However, future occurrences of certain queries may follow a predictable pattern, and we can use this prediction to build histograms "proactively". We will explain how these proactive histograms can be obtained and used together with "reactive" histograms, and show some interesting experimental results, especially when only a small portion of the queries are predictable.