Why couldn't it be serious? Your database already has your data so you don't have to stream it to another machine for training. If you have a developed model th

Why couldn't it be serious? Your database already has your data so you don't have to stream it to another machine for training. If you have a developed model that needs retraining on new data, save yourself the network bandwidth (and time!) and just make the DB retrain. This doesn't replace the exploration phase of model-building but if you already know gradient descent works for you then what's wrong with this approach?