Effective Generation of Relational Schema from Multi-Model Data with Reinforcement Learning

Abstract

To handle data variety in one project, some researchers proposed using multiple databases or one multi-model database to manage various data. However, considering that the predominated Relational Database Management Systems (RDBMSs) in the current market have powerful capabilities such as query optimization and transaction management, we propose using an RDBMS as a unified platform to store and query multi-model data. But the mismatch between the complexity of multi-model data structure and the simplicity of flat relational tables imposes a grand challenge. To address this challenge, we adopt the reinforcement learning method to design a workload-aware approach that could directly learn a relational schema to store multi-model data by interacting with an RDBMS with the given queries and data. To choose the right actions in the learning process, we propose a variant Q-learning algorithm (Double Q-tables) along with functions for updating the tables, which could reduce the dimension of the original Q-table and improve learning efficiency. Experimental results show that our approach could generate a relational schema with superior performance in terms of query response time and storage space cost over a multi-model storage schema.