面向时空数据场景的数据库索引选择框架  

A Framework for Database Index Selection over Spatio-temporal Data Scenarios

作　　者：徐康镭 乔少杰 陈金勇[2] 张桃 高林 冉黎琼 谢添丞 于泳 彭钰寒 焦育威 XU Kanglei;QIAO Shaojie;CHEN Jinyong;ZHANG Tao;GAO Lin;RAN Liqiong;XIE Tiancheng;YU Yong;PENG Yuhan;JIAO Yuwei(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China;The 54th Research Institute of CETC,Shijiazhuang 050081,China;Faculty of Artificial Intelligence and Big Data,Yibin University,Yibin 644000,China;Sichuan Digital Transportation Technolgy Co.,Ltd.,Chengdu 610225,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]中国电子科技集团公司第五十四研究所,河北石家庄050081 [3]宜宾学院人工智能与大数据学部,四川宜宾644000 [4]四川数字交通科技股份有限公司,四川成都610225

出　　处：《无线电通信技术》2023年第4期665-673,共9页Radio Communications Technology

基　　金：国家自然科学基金(62272066,61962006);四川省科技计划(2021JDJQ0021,2022YFG0186,2022NSFSC0511,2023YFG0027,2022YFG0325,2021YFG0029);教育部人文社会科学研究规划基金(22YJAZH088);宜宾市引进高层次人才项目(2022YG02);成都市“揭榜挂帅”科技项目(2022-JB00-00002-GX,2021-JB00-00025-GX);成都市重大科技创新项目(2021-YF08-00156-GX);中国电子科技集团公司第五十四研究所高校合作课题(SKX212010057);四川省教育厅人文社科重点研究基地四川网络文化研究中心资助科研项目(WLWH22-1);成都信息工程大学科技创新能力提升计划(KYTD202222);成都海关科研项目(2022CK008)。

摘　　要：索引是一种用于提高数据库查询效率的数据结构,良好执行计划的生成和选择很大程度上取决于数据表上是否存在合适的索引。然而,面对复杂的时空数据场景,现有的索引选择方法存在很多不足,如无法高效处理大量范围查询、容易造成索引冗余、无法有效应对动态的工作负载等问题。针对上述问题,提出一种新型基于深度确定策略梯度(Deep Deterministic Policy Gradient,DDPG)模型和索引效益评估网络的索引选择框架:ST-IS(Index Selection method tailored for Spatio-Temporal data scenarios)。ST-IS集成了对SQL语句和索引集的新型编码方法,有效避免索引冗余或索引缺失。ST-IS使用索引效益评估网络来优化强化学习的奖励机制,同时支持动态的工作负载。大量实验结果表明,在真实时空数据集和TPC-DS标准数据集上,ST-IS的性能优于现有的索引选择方法。The index is a data structure used to improve the efficiency of database queries,and the generation and selection of a good execution plan depends much on the presence of appropriate indexes on data tables.However,existing index selection methods have many shortcomings when facing with complex spatio-temporal data scenarios,such as inefficient handling of a great number of range queries,potential index redundancy,and inability to effectively handle dynamic workloads.Aiming to cope with these problems,a new index selection framework called ST-IS(Index Selection method tailored for Spatio-Temporal data scenarios)is proposed,which is based on the Deep Deterministic Policy Gradient(DDPG)and an index utility evaluation network.ST-IS integrates a novel encoding method for SQL statements and index sets,effectively avoiding index redundancy or missing indexes.ST-IS uses an index utility evaluation network to optimize the reinforcement learning reward mechanism,while supporting dynamic workloads.Extensive experiments are conducted and the results show that ST-IS outperforms existing index selection methods on real spatio-temporal datasets and the TPC-DS benchmark dataset.

关 键 词：时空数据 强化学习 索引选择 查询优化 移动数据挖掘 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]