One area in which systems for machine learning are wanting (TensorFlow, PyTorch) is in their support for Big Models and Big Data. In contrast to modern relational systems which scale to large data sizes and multiple machines quite well out of the box, getting machine learning computations to work in a distributed setting or with large models is often very challenging. In this talk, I argue that the fundamental problem is lack of abstraction in these systems. I argue that it makes sense to re-design these systems from the ground up, applying many of the lessons from the heyday of relational database system design in the 1970’s and 80’s.