After attending tonights got:ruby I became obsessed with and started installing rvm. It worked swimmingly.
I’m looking forward to getting my development machine off of the embarrassingly low version of Ruby and experimenting on the latest versions without (unintentionally) screwing up my production environment.
Indeed after this week I’ve got a lot to google, lookup, and research. I’ve decided I have got to get good at JQuery—probably using it on a small Rails 3 project in the near future—and learned that Jasmine has it all over ScrewUnit. This may mess up my Haskell plans—not that I’ve ever messed that up before.
After that we played a lot of ping-pong, then I decided to clean up my hard drive, because it was really time. Then I realized if I’m gonna do this project as Rails3, I better make sure I can deploy it on dreamhost with bundler. If this sounds like procrastinating it’s cause it is.
I also checked email, read Anders’ blog on the Elabs site, and watched some of Germany beating Ghana. None of this is writing the blog post and that’s because today’s blog post is going to be hard.
Let’s get this out of the way first. Today was another great day at ELabs. With my Jet Lag finally kicked I really was able to contribute, using a couple ideas we’ve had before on 8th Light projects to get some cool work done on the project I’ve been working on this week.
I even wrestled for the keyboard a bit (although they have two keyboards - so it was lonely, I wrestled with Pixie) then we topped it all off a cool Ruby user group meeting.
Since I’m awake, the ELabs guys have to suffer my jokes like the 8th Light guys do. Now they know why they sent me out of the country.
What’s been really nagging at me is the way that the ELabs boys work. See, they are what Brian Marick would call “classicist” TDDers—they essentially never use mocks. I on the other hand generally mock out a lot.
What Jonas said to me was, and I’m going from memory here, “I’d rather the tests run slow than be right.” Of course taken as stated, this is a truism. Nobody is voting for slow tests, although I might argue that the tests that verify correctness are often different than unit tests, and that unit tests are verifying a different thing.
Aside from speed mocks provide value in designing apis by letting you design it with a client using it, before the object itself exists, and in thinking of systems as objects passing around messages you can really focus on their interactions instead of the implementation details. Mocks lead to a cleaner design all told…except…
Except I usually stub out things for speed. Objects I’m stubbing are webservices, database objects, file access, etc. And in the best cases the benefit I get is speed, and in the worst cases I get a tangled mess of mocks where my code and tests are coupled, but nothing necessarily works. I’ve completely traded speed for correctness.
Why is that? Why am I not getting all the purported benefits of mocking. In a related note on the way here I started reading “growing object oriented software, guided by tests” by Steve Freeman and Nat Pryce.
In a throw away comment they say, “Never mock out third party APIs, write thin classes that wrap those with integration tests, and mock out your objects interaction with those classes.” Now maybe you think you haven’t done this, but let me give you some typical examples of mocks:
Well look at that—those are two third party APIs aren’t they? And they are often mocked out everywhere! Trading speed for correctness again. “Okay fancy boy”, you’re saying—and really why are you calling me that, “Then we shouldn’t use mocks. Problem solved.” Not so fast. Have you ever had a problem where you're testing that an object is saved and you do this?
But then you think, “Gosh I really don’t care how they save the values, I
just care that it gets saved.” Right now you have to call
update_attributes to make this pass, but you could call
update_attribute, or set the object and save!, or hell use
SQL although you shouldn’t.
This makes for a crappy test—one you’re likely to break then fix later instead of changing in the deliberate process of TDD.
What if instead you did this:
Of course the other things are still true—assuming it’s an active record object, but maybe it’s not. Maybe it’s a wrapper around it, or maybe it’s datamapper object, I don’t really don’t care.
I’ve gone ahead and mocked this call because that’s the call I really want to see—and somewhere else I’ve test driven that call against the database. Now let’s imagine you’ve got more complicated scenarios. Ever done this?
This is a simple example—but I’m sure you’ve seen worse. The problem is again that you really need one method, maybe two, but you probably figured out in your head the algorithm beforehand, then test drove it. Maybe you know if object2 was true you needed to set unrelated object3’s flag—so you started writing those tests.
Let’s look at that rule again though—don’t mock a third party API. Well what defines a third party API? Is it a third party if another company wrote it? How about the same company—but on a different project? Or the same project but a long time ago? Or YOU six weeks ago and you forgot it.
Isn’t it a third party API again? What’s the difference? Should
you then mock out
objectnew—and test drive it with real objects?
Because it’s pretty clear the API you want doesn’t exist yet,
and that’s the crux of my long wrambling argument that I’ll probably better
express and codify when it’s not late and I’m not tired:
Stop mocking the API you already have—mock the API you need
And then if by chance it matches the API you already have—by all means just substitute the real object in. But write the implementation with mocks first—without using any real API unless it matches coincidentally.
Pretend the object doesn’t exist yet, and then mock. And maybe then you’ll get what you’re looking for.