Hybrid Indexes by Exploring Traditional B-Tree and Linear Regression

Abstract

Recently, people begin to think that database can be augmented with machine learning. A recent study showed that deep learning could be used to model index structures. Such learning approach assumes that there is some particular data distribution in the database. However, we argue that the data distribution in the database may not follow a specific pattern in the real world and the learning models are usually too complicated, which makes the training process expensive. In this paper, we show that linear models can achieve the same precision as models trained by deep learning using a hybrid method and are easier to maintain. Based on this, we propose a hybrid method by exploring traditional b-tree and linear regression. The hybrid method retrieves data and checks whether the data can benefit from learning approach. We have implemented a prototype hybrid indexes in Postgres. By comparing with b-tree, we show that our method is more efficient on index construction, insertion, and query execution.