Anatomy of Learned Database Tuning with Bayesian Optimization

Abstract

Database Management System (DBMS) tuning is central to the performance of the end-to-end database system. DBMSs are typically characterised by hundreds of configuration knobs that impact various facets of their behavior and planning abilities. Tuning such a system is a prohibitively-challenging task due to the obfuscated knob inter-dependencies and the intimidating size of the design space. The general vendor recommendation is to sequentially tune each knob, which further exacerbates the time-consuming nature of the task. To overcome this, recent work in the realm of self-driving database systems proxy the design problem through Machine Learning. Among the most prominent proxies in self-managing databases literature is the Bayesian-inference proxy. The purpose of this proxy, or surrogate in Bayesian Optimisation parlance, is to learn the inter-knob relationships and how they relate to the overall performance, independent of any human guidance. To this end, one of the goals of this work is to shed light on the common design patterns we identify in Bayesian-driven DBMS tuning agents. Second of all, we aim to provide a handbook for implementing such agents through the lens of a new tuning framework that leverages a multi-regression proxy.