A Query-Level Distributed Database Tuning System with Machine Learning

Abstract

Knob tuning is important to improve the performance of database management system. However, the traditional manual tuning method by DBA is time-consuming and error-prone, and can not meet the requirements of different database instances. In recent years, the research on automatic knob tuning using machine learning algorithm has gradually sprung up, but most of them only support workload-level knob tuning, and the studies on query-level tuning is still in the initial stage. Furthermore, few works are focus on the knob tuning for distributed database. In this paper, we propose a query-level tuning system for distribute database with the machine learning method. This system can efficiently recommend knobs according to the feature of the query. We deployed our techniques onto CockroachDB, a distribute database, and experimental results show that our system achieves higher performance under typical OLAP workload. For all categories of queries, our system reduces the latency by 9.2% on average, and for some categories of queries, this system reduces the latency by more than 60%.