Feature generation pipeline for machine learning

A method comprising:

• executing a plurality of jobs, each of which computes a different set of feature values for a different feature of a plurality of features associated with videos;
• storing a feature registry that lists each feature of the plurality of features;
• after executing the plurality of jobs and storing the feature registry, receiving a model specification that indicates a set of features for a model;
• for each feature in a subset of the set of features:
• identifying a location in storage where a value for said each feature is found, and
• retrieving the value for said each feature from the location;
• creating a feature vector that comprises, for each feature in the set of features, the value that corresponds to said each feature;
• using the feature vector to train the model or as input to the model that outputs a score relative to a particular user-video pair;
• wherein the method is performed by one or more computing devices.