View-Size Estimation in Self-Organizing Databases

Data-intensive systems routinely use derived data, such as indexes or materialized views, to improve query-evaluation performance. In this context, the problem of dsigning derived data is as follows: Given a database and a set of queries, return definitions of derived data that, when precomputed and stored in the database, would reduce the evaluation costs of the queries. Designing materialized views and indexes is an important part of automated query-performance tuning in data-management systems that experience changes over time, where a system addresses the performance requirements of current frequent and important queries by periodically reconsidering and rematerializing the stored derived data. In this paper we focus on the accuracy of estimating the sizes of views that are considered for materialization. We analyze and experimentally evaluate several size-estimation techniques and suggest guidelines for using the techniques depending on system parameters. We describe experimental results obtained in our prototype self-organizing database system QPET. Our guidelines are also applicable to estimating the sizes of intermediate results in query optimization.