From T261403#6541529:
@MGerlach, maybe it would be easier if we just have a single repository, and then use multiple requirements.txt files so that the code we ship in the production query service doesn't have all of the heavier libraries used for training the model, and the code used for training the model doesn't have the HTTP API libraries, etc? The advantage would be reduced overhead in making updates to code shared across training / querying (e.g. you wouldn't have to update a library, commit and push, then update the training and query repos to use the updated library).