Research On Method Of Database Index Selection Based On Reinforcement Learning

The database system has been operating maturely and stably in the industry,and the era of big data has put forward higher requirements for the database system.The increase in data volume requires the database system to have faster query speed and higher system throughput.The type pattern of data is increasing,making the workload fast and diverse,requiring the database system to have the ability to quickly and accurately respond to dynamic changes in workload.As the size of data continues to increase,database performance tuning has become increasingly important to researchers,enterprises,and database administrators.The index selection is one of the important issues in database performance tuning.Setting appropriate indexes in the system can speed up the data access process,and using inappropriate indexes wastes resources and causes unnecessary disk overhead and index maintenance costs to the system.Although some research progress has been made on the index selection methods based on reinforcement learning,the related work does not take into account the impact of index interaction,and the query performance of the database does not improve significantly with the increase of the number of indexes and even decreases.Most of the index selection methods based on reinforcement learning are only able to select singlecolumn indexes,which misses multi-column indexes that can significantly improve database query performance,such as cover index.Therefore,this thesis studies the index selection method based on reinforcement learning,and the main work is as follows:1.To address the impact of index interaction on index selection,this thesis defines the index selection problem as a Markov decision process,and by defining the environment state,the action space of the intelligence and the reward function in the deep reinforcement learning process,the database state after index selection by the agent is compared with the query cost in the initial state and the database state after the last selection,respectively,fully considering the index.The possible interactions between them are fully considered,and the optimal index combination can be selected.The experimental results demonstrate that this method can improve the query performance of the database to a greater extent than the index combinations selected by the current classical index selection methods.2.In this thesis,we study the DDQN algorithm with excellent performance in deep reinforcement learning,and use the dynamic exploration strategy proposed in this thesis as the exploration strategy of DDQN,which balances the relationship between exploration and exploitation and enables the agent to learn the optimal strategy;for the priority experience replay makes the training samples concentrated in a smaller subset,which may lead to the network model overfitting and falling into local optimum.To solve this problem,a uniform sampling-based priority replay method is designed to combine uniform sampling and priority selection,which increases the diversity of training samples.3.For the problem that only single-column indexes can be selected,this thesis generates candidate indexes by heuristic rules,which can realize the simultaneous selection of single-column indexes and multi-column indexes.Firstly,the characteristics of the workload are analyzed,and five heuristic rules are proposed to limit the number of candidate indexes and provide high-quality candidate indexes,while reducing the state space and action space dimensions in deep reinforcement learning to speed up the training process of index selection agent.Secondly,this thesis proposes an index selection method based on heuristic rules and deep reinforcement learning.Finally,the results are validated by comparison experiments,which show that the indexes selected by the proposed method in this thesis can improve the query performance of the database more significantly compared with the existing methods.