In the practices of machine learning, Feature Engineering is a crucial step that converts raw data to the inputs of models. This process conventionally relies on data processing languages (typically SQL) and now sees arising challenges from the advancement of machine learning techniques. We present the design of F3, a DSL and a compiler developed at Meta. We show how F3 transforms the inspirations from functional programming and type theory to an industrial grade software architecture that empowers a platform that serves billions of users.