Explaining unexpected performance in database query execution
User satisfaction when considering query performance of a database management system (DBMS) is a crucial component to enabling broader DBMS adoption. Stability and predictability of query performance that imply that similar query inputs should have similar execution performance, are major goals for industrial vendors toward respecting service level agreements (SLA). This is amplified in modern decision support systems that are characterised by parametrised queries and user defined functions (UDF), which frequently repeat the same query template with different parameter values.
The core of the problem of unsatisfied users come when query performance changes suddenly. For instance, a query might complete in seconds one day, and take hours the next day, while the only change to the query might be a different choice of parameters. In this project, we examine the source of unexpected performance, providing users with a clue of why performance has suddenly changed. The project combines database technology with machine learning (ML) techniques in order to predict query performance and isolate and explain outliers.
This project is done in collaboration with Microsoft Research (USA) and industry.
Leader: Renata Borovica-Gajic
Collaborators: Microsoft Research USA
Computing and Information Systems
Networks and data in society
data structures; database systems; machine learning