Cloud Database Tuning with Reinforcement Learning

Chenxia Han, Chaokun Chang

Keywords: Reinforcement Learning, Database Tuning

TL;DR: A scalable database tuning system in reinforcement approach

Abstract: In database management systems (DBMSs), especially cloud DBMSs, configuration tuning is one of the key factors that influence database performance. For a long period of time, the tuning job of databases is done by experienced database administrators (DBAs), which is time-consuming and sub-optimal. Recently, with the development of machine learning, automatic tuning tool starts to play a significant role in DBMSs. Among all the learning-based methods, reinforcement learning has the greatest potentiality to find the optimal or near-optimal configuration. There are already some reinforcement learning applications in database tuning, nonetheless, none of them provide fully executable code. In this project, we implement an independent version of auto-tuner to re-produce the current works on MySQL. Based on a careful design of the system, our code can be easily applied to other databases like JDB, LevelDB, etc. We believe it will benefit a lot to those database administrators who know a lot about databases but lack the knowledge of reinforcement learning. The code is available at https://github.com/ChaokunChang/GDBTuner. The 5 minutes video can be found at Google Drive https://drive.google.com/drive/folders/1Z1MvtBnQ522zPElNu-VSrAMvKcX523LE?usp=sharing.