We introduce Semblance, a machine learning feature generation system for both model training and real time prediction. Semblance is based on the foundation of Functional Reactive Programming, enabling a programming model which guarantees that features are correct by construction with regard to time ordered events. Semblance features are platform agnostic, which allows the same feature definition to be evaluated in different platforms, such as Scalding, Spark, Flink, Storm, or other similar map-reduce style or stream processing systems. In addition, Semblance enables a lambda architecture system, which we have implemented in production to evaluate features in real time within a few milliseconds.
