Also, CSV can’t easily handle nested objects. If the data model is even slightly more complex than a plain table, it doesn’t make much sense. I’d also argue that even if the source data is stored in an RDMS without exotic data types, a JSON with a nested object representation is probably going to be more friendly even to non-developers than multiple files with opaque foreign keys linking back and forth.