The short answer is that the "relational" model is a generic logical theory for managing data under a closed world assumption. "Non relational" mostly is about saying "I don't need a logical model, I need a persistence framework", wherein you are just looking to implement durable storage, but aren't really managing data in a logical manner.
Some network/graph databases are arguably a special case of the relational model, but they all vary to the extent under which they enforce logical integrity vs. exposing the innards to muck with (same with all RDBMS' frankly).
William Kent's 1978 book _Data and Reality_ is a great read to think of data modeling and integrity issues generically, irrespective of whether you use an RDBMS or a non-relational database. He highlights certain logical concepts as universal regardless of whether you use tables, K/V pairs, or graphs of binary relationships.
But this is the story of an engineer using a profiling tool to solve a performance problem, and the solution was to drop resolution in his probabilistic data model. I don't see how BCNF would have avoided this if the system were implemented in RDBMS. The actual solution reminds me of the quantization step in JPEG processing - its lossy of the high-entropy but imperceptible data.
To your question: cost accounting - technology neutral, and brutal in it's power to decide.