At Netflix, we are continually looking to improve our member recommendations by following a data driven approach based on machine learning algorithms. To enable faster iterations with high confidence, there are two key components: collecting historical fact data, and providing tools to do feature generation and model training easily. We describe the components needed for building a time machine for feature generation using Apache Spark that enables our researchers to easily try ideas using historical data and makes running offline experiments and transitioning to A/B tests seamless.