Machine learning in SQL by translation to TensorFlow

ABSTRACTWe present sql4ml, a framework for expressing machine learning (ML) algorithms in a relational database management system (RDBMS). The user writes the objective function of an ML model as a SQL query, then sql4ml translates the query into an equivalent TensorFlow (TF) graph, which can be automatically differentiated and optimized to learn the model weights. Sql4ml makes the database a unified programming environment for feature engineering, learning/inference, and evaluating models. The proposed approach is more expressive than using ready-made ML algorithms, but abstracts away the details of the training process. We present the architecture of sql4ml and describe the method for translating an objective function in SQL to a TensorFlow representation. We show how recent ideas from Factorized ML [7] can be leveraged to efficiently move data between a database and an ML framework. Finally, we present experimental results regarding both the proposed translation and the optimization techniques for data transfer. Our results show that translation time is negligible compared to time for data processing, and that the optimization techniques achieve up to 50% improvement in the export runtime and up to 85% decrease in the size of the exported data.