Doing Machine Learning the Uber Way: Five Lessons From the First Three Years of Michelangelo