[gedrap]

This. The problem is that in some cases, the legacy system is really messed up and making it unit testable is a big refactoring task on it's own.

In these cases, I try to write some functional tests first (such as calling restful endpoints and checking the response, or using a headless browser). Not great but much better than nothing. Any ideas on how to do it better?

[blt]

The closest thing I've seen to an actual definition of "legacy code" is "code that's very difficult to unit test." It's a shitty catch-22.

[philbarr]

Feathers defines legacy code as code without accompanying tests in "Working Effectively with Legacy Code". Even to the point where someone he knows said of a particular shop, "they're writing legacy code, man!"

[crdoconnor]

Actually, I typically surround legacy code with functional tests instead. Unit testing is only really useful on blocks of code that you can safely wall off from the rest of the code base. Big balls of mud by their very nature don't really have that.

Much of the necessary refactoring for legacy code involves decoupling, which inevitably means changing method signatures and even replacing entire methods. If you surrounded those methods with unit tests which will break even when the functionality doesn't, you've made the code more resistant to refactoring, not less.