Database Native Approximate Query Processing Based on Machine-Learning.

With the worldwide digital transformation, many databases with large volumes appear and provide interesting insights analyzed by data scientists through all kinds of tools. The large volumes of them inevitably increase the workload of calculation, lengthen the response time of applications and negatively impact the user experience. Approximate Query Processing (AQP) is proposed to alleviate this issue. Although many researchers continuously improve the performance of AQP with the help of Machine Learning, there are few studies on embedding Machine Learning based AQP inside the relational database through User Defined Functions (UDF). In this paper, we focus on one specific kind of aggregate queries and present two different implementations to embed one Machine Learning based AQP inside Relational Database Management System (RDBMS) by taking advantage of UDF. Both implementations are able to calculate estimates with acceptable errors, and the implementation with external training and internal query processing has even better performance in term of response times.