MLearn: A Declarative Machine Learning Language for Database Systems

This paper outlines the requirements of our ML2SQL compiler that allows a dedicated machine learning language (MLearn) to be run on different target architectures. The language was designed to cover an end-to-end machine learning process, including initial data curation, with the focus on moving computations inside the core of database systems. To move computations to the data, we explain the architecture of a compiler that translates into target specific user-defined-functions for the PostgreSQL and HyPer database systems. For computations inside user-defined-functions, we explain the necessary tensor datatypes and the corresponding functions. We base the explanations on an accompanying example of linear regression. To face the challenges to database systems arising from array-like data, we propose such solutions as integrating ArrayQL as stored procedures to unify the relational and array perspectives.