关键词: Learned Index, B-Tree, 可更新的自适应学习索引, 在线分析处理数据库, 静态数据, 优化调整

Abstract:

Learned Index is an index which establishes the relationship between input data and data storage location through training model. It can learn the information of data distribution， but different data distribution will affect the balance between model accuracy and model complexity. In order to explore the applicable scenarios of Learned Index， the performance of it and An Updatable Adaptive Learned Index （ALEX） were tested by using data of different distribution and amount， and compared with B-Tree. It is found that the time to build Learned Index of big data is shorter than that of B-Tree， and it has obvious advantage in reading operation performance and storage space size， but its writing operation performance is poor. Therefore， it is concluded that Learned Index is more suitable for OnLine Analytical Processing （OLAP） databases in big data scenarios， which is used for static data storage and query operation. Based on the structure of B-Tree， Learned Index was optimized and adjusted， and finally the performance of the optimized Learned Index was significantly improved in the read and write operations of large volume data， which beats B-Tree by up to 2 times on reading operation performance and 3 times on writing operation performance.

Key words: Learned Index, B-Tree, An Updatable Adaptive Learned Index (ALEX), OnLine Analytical Processing (OLAP) database, static data, optimization and adjustment