I have generally found NoSQL to be a disaster. Like agile processes, it allows you to dispense with certain disciplines, but for use at scale and over time it requires you to engage in substitute disciplines. Too often these are not practiced. From a recent work chat with minor adaptation:

Data hygiene is crucial. I wouldn’t be opposed to broader NoSQL/JSON use if we used JSON schemas wherever appropriate, but at that point it is probably simpler to flatten the data into tables.

A good schema is a species of defensive code; e.g., you can have higher confidence that the value you are reaching for is actually there no matter how old the document.

See also: Complex