Reproducibility is important throughout the entire data science process. As recent studies have shown, subconscious biases in the exploratory analysis phase of a project can have vast repercussions over final conclusions. The problems with managing the deployment and life-cycle of models in production are vast and varied, and often reproducibility stops at the level of the individual analyst. Though R has best in class support for reproducible research, with tools like KnitR to packrat, they are limited in their scope. In this talk we present a solution we have developed at Domino, which allows for every model in production to have full reproducibility from EDA to the training run and exact datasets which were used to generate. We discuss how we leverage Docker as our reproducibility engine, and how this allows us to provide the irrefutable provenance of a model.
