Adaptive query processing has emerged as an attractive solution to dynamically changing runtime environments, and unknown data distributions in many scenarios. These scenarios involve querying over web sources where little information is usually available about the data sources, continuous query processing where the data characteristics change frequently, and even single site database systems where in many cases, only unreliable statistics are available. Probably the most aggressive of the adaptive query processing techniques proposed is eddies, which continuously reoptimizes a query by changing the order in which operators are applied to tuples on a per-tuple basis. We observe that, even though the eddy can, in principle, choose to route each tuple in a different order, the query execution plans it can affect for multi-join queries can be significantly constrained by the routing decisions it made early during the query execution. In this talk, I will present a modified eddy architecture, called STAIRs, that allows the eddy to manipulate the state stored inside operators, and thereby lift the burden of routing history. I will also discuss our implementation of eddies in the PostgreSQL data management system, and show that the benefits of such adaptivity can be had at a very low cost. Finally, I will briefly discuss some of my recent work on exploiting correlated attributes in sensor network query processing.

Bio:

Amol Deshpande is a PhD candidate in the Computer Science Department at University of California at Berkeley, and plans to graduate in August 2004. He works in the Database Research Group with Prof Joseph M. Hellerstein, and his main research interests include query optimization and query processing in a variety of scenarios.