Kind of, yes. If you are going to be doing this at scale, you will either need to pin servers to shards of users, or retrieve the DB from some durable storage "just in time", or... use a network mount. Note also, that with small-enough databases you can do the same thing Apple does - fetch your entire tenant from the "big datastore", set a lock on it, do the thing that you want to do - and then write it back into the "big datastore" at the end.

I guess if the tenant isolation is more important than the disk/network io overhead this would introduce then it's viable solution.

It'd probably work better if tenant == user and more specifically tenant == single-user-session.