Consider making it work with online learning algorithms

[johann-petrak]

This is interesting for huge datasets and stream learning. The LF would essentially just get document by document and pass extracted features on to the online learning algorithm (e.g. Vowpal Wabbit, AROW). This could happen via streaming/reading from a pipe or via REST.

The learning algorithm must be able to write an updated model at regular intervals or when somehow triggered from the outside.

The application PR could also support a trigger to reload its model, thus supporting using the very latest model from e.g. a stream-learning job.