Well technically you don't join (document store approach) or you keep your IDs and FK's as surrogates outside the JSON blob and use them as indexes only (hyprid mongo-ish store).
What it does give you is a JSON object store with somewhat better stability and real transactional guarantees compared to say CouchDB or MongoDB.
However, the issue I have with this approach is more that there is some opportunity to add more processing into the database black box, which is expensive to scale up. I'd rather it wasn't type aware of JSON (i.e. it was just a varchar(max) column) and all that processing sat in the application itself which is cheap to scale out.
What it does give you is a JSON object store with somewhat better stability and real transactional guarantees compared to say CouchDB or MongoDB.
However, the issue I have with this approach is more that there is some opportunity to add more processing into the database black box, which is expensive to scale up. I'd rather it wasn't type aware of JSON (i.e. it was just a varchar(max) column) and all that processing sat in the application itself which is cheap to scale out.