MS Thesis Defense: Extendable Rule-Based Action Generation for Self-Driving Database Systems (Mike Xu)
Database management systems (DBMSs) have become more complex to meet increasingly demanding usage. To owners and operators, the need for a self-driving DBMS that can automatically tune and optimize itself without human intervention is apparent now more than ever. Such a self-driving DBMS considers a set of candidate actions to apply to reach a configuration that improves performance for a given workload. Furthermore, the DBMS would continuously adjust the configuration in anticipation of changing workloads and data distributions.
Efforts to architect self-driving DBMSs suffer from the engineering overhead of combining different tuning subtasks, such as index tuning and knob tuning. These individual subtasks have a vast candidate action space, requiring the tuning algorithms to reduce the action space before searching. However, each subtask and algorithm has its own representation of the action space and methods for obtaining requisite inputs for the algorithm, including a representative workload and the database schema.
This thesis presents an extensible framework for defining the action space of tuning subtasks. The framework allows the self-driving DBMS engineer to define action types, rules for constructing the action space, and input information for the tuning algorithm – all in a standard interface shared across subtasks. The framework reduces the overhead of developing and evaluating new tuning algorithms and allows the DBMS to dynamically define the subset of the action space to explore in any given tuning task. By restricting the search space before executing the tuning algorithm, the framework reduces the time expended on evaluating suboptimal configurations. It thereby improves the speed at which the algorithm converges on a solution.
We then use this framework to alter and restrict the action spaces of existing algorithms for index tuning and knob tuning. We demonstrate that by filtering the search space against low-quality candidate actions, the framework enables tuning algorithms to converge more quickly on tuning actions that can match or outperform the baseline solution.
- Andy Pavlo (Chair)
- Zhihao Jia
More Info: https://www.cs.cmu.edu/calendar/163737037