ATestDoubleDilemma

I'm conflicted and confused about test doubles. Here's a brief history of my life with with them:

• Began TDD without any test doubles. Was buried under complexity within a month.
• Started using Michael Foord's "mock" library for Python. Liked the library, but found it tedious to create so many mocks manually.
• Forked "mock" to create Dingus, which automatically replaces all objects in a class's namespace with recording, super-permissive stubs.
• Found that, now that all code is 100% isolated, it's almost impossible to refactor across classes with confidence unless I totally rewrite them.

There are wonderful things about enforcing complete class-level isolation: extremely fast execution (typically 1ms per test with Dingus); easy to test legacy classes without their dependencies; no penalty for "turtles all the way down"/"wishful thinking" style of design. But it seems that we pay for this with much more expensive refactoring.

After reflecting on all of this, I asked a bunch of experienced TDDers for advice and basically got every answer in the spectrum from "I double everything" to "I almost never use doubles." So, I'm confused. How can there be so little consensus on something so important? I feel like I must be missing something!

Gary Bernhardt