have to disagree with this. By forcing yourself to work with a data layer so abstracted that you can't even reference whether you're dealing with a JSON document or a set of twelve joinable tables, you're going to write the most tortured and inefficient application. Non trivial applications require leaky abstractions.