What is the difference between a mock, a stub, and a fake, and when do you use a real dependency instead?

A stub returns canned responses so the code under test can proceed. A mock also records interactions so you can assert they happened, verifying behavior rather than state. A fake is a lightweight working implementation, like an in memory repository. Use real dependencies when the integration itself is the risk, and test doubles when the dependency is slow, costly, or hard to force into a specific state.

A stub is passive, it just returns whatever you told it to return so the code can keep going. A mock is a stub that also remembers what happened to it, so you can assert the payment service was called exactly once with these arguments. A fake is a real but simplified implementation, like an in memory version of a repository, which behaves properly but is not production grade. The distinction that matters in practice is that stubs and fakes support state verification, asserting on the outcome, while mocks support interaction verification, asserting on the conversation. I lean toward the former, because asserting on interactions couples the test to implementation details, and then a refactor breaks a hundred tests without a single behavior changing. Where I use real dependencies is anywhere the integration is the risk. If the question is whether our query actually works against Postgres, a mocked database answers nothing, so I run the real thing in a container. That is my default for the integration layer now, since containers made it cheap. I mock third party services I do not control, then back that with scheduled contract tests against their sandbox.