When you first start testing a newly created application, fixtures might seem very appealing. They're easy to write, they make sense, and they quickly create valid or invalid instances for you to test with. Unfortunately, fixtures don't scale with a growing project. They quickly get out of hand, and you'll end up spending more time fixing your fixtures than your tests and code. So fuck you fixtures, and good riddance.
Fixtures are the devil because:
Rails 2 fixtures are nicer, but it's much of the same. If you don't believe me, just keep using them and you'll know what I'm talking about at some point.
One interesting case my friend came across the other day was using fixtures for rspec tests. As described above, I think model fixtures for testing is both brittle and hard to maintain. However, he was running across the case:
it "should save to the freakin database" do @some_model.save.should == true debugger end
This test passes with flying colors... But seeing is believing, and if you login to the test database, you won't be seeing a saved record. This is generally a good thing because once the test is finished, the transaction is rolled back, and you'll have clean data independence between tests. The downside is when you're testing something like Sphinx that assumes there's stuff in the database to work with, it won't work. The above example will pass, and if you break at the debugger line, @some_model.new_record? will be false.
Other than that valid case, most problems should be solvable by stubbing and mocking. If not, then refactor your code so that it is! One particularly delicious piece of syntactic sugar called Factory Girl. For a short description of what it does, check out this blog post.