This is one of those things that I would hope goes without saying, but obviously doesn't... Always test against what you expect to see in production. If you test against something else first (in this case mocking through an in-memory DB) to make the testing of other parts faster/easier then that is fine, but once those tests are done you still need to do a final full test against the real stack(s) you expect to see in production.

On the mocking thing: I thought the point of that was to completely avoid DB calls for speed when testing other logic layers, so even a fast in-memory DB isn't needed and anything (i.e. just some code that manufactures a response object) that produces a result set in the right format will do? In that case even using an in-memory DB is less efficient than the other option so is at best a "half optimisation". Am I missing a point here?