DBMS Fitting

Deep Neural Networks (DNNs) have successfully been used to replace classical DBMS components such as indexes or query optimizers with learned counterparts. However, commercial vendors are still hesitating to put DNNs into their DBMS stack since these models not only lack explainability but also have other significant downsides such as the requirement for high amounts of training data resulting from the need to learn all behavior from data.

In this project, we investigate alternative approaches that incorporate domain knowledge to obtain more reliable learned DBMS components requiring less training data. While the high-level design of the DBMS component is still specified by code, we optimize it for a particular workload and hardware using differentiable programming. Differentiable programming is a recent shift in machine learning away from the direction taken by DNNs towards simpler models that take advantage of the problem structure. We successfully applied this technique to learned indexing and cost modeling for query optimization.