[MartinCron]

In my experience, it's all about layers.

I do both, I have a set of integration tests create a "safe copy" of the database and populate it with seed data, so I can trust that my data-layer stuff works correctly, while my pure unit tests interact with a super-fast fake database implementation.

I actually wrote the in-memory fake db before I wired up a real database. It made the early iterations much faster without having to make changes to multiple layers. I could have not bothered with writing tests against a "safe copy" of the database, but I sleep better knowing that the assumptions about how the persistence types work in my unit tests are validated by my integration/data tests.