Practical decision systems require much more than end-to-end learned models. This talk will focus on research and engineering questions on machine learning robustness and the two broad categories of work in that area.