RT @josh_tobin_
In the real world, data distributions are almost never static.

So if you want to use ML in production, your goal should be to build a continual learning machine, not a static model.