Regression-Based Self-Tuning Modeling of Smooth User-Defined Function Costs for an Object-Relational Database Management System Query Optimizer

We present a new approach to modeling the execution costs of user-defined functions (UDFs) for the query optimizer of an object-relational DBMS (ORDBMS). Our approach self-tunes a cost model incrementally based on the costs of the recent executions of a UDF. The approach is centered on a feedback loop in which the feedback information comprises individual UDF execution records. Each execution record contains the cost variable values used in the execution and the resulting CPU and disk I/O costs. This feedback information is saved in the execution log and used in a batch to update the cost model. Furthermore, our approach handles nominal cost variables by maintaining separate cost models for recently used values of the variables. We have built a framework that implements the feedback loop in a commercial ORDBMS. Then, we have performed experiments using common database UDFs with smooth cost variations and incrementally modeling the data using multiple regression. The experimental results demonstrate the adaptive accuracy that makes the cost model stabilize quickly while incurring small errors in cost estimations. Our approach has the advantages of incurring little overhead while tuning the cost model continuously throughout the UDF executions.