Tuesday, 1 March 2011

Mock Roles not Types

"if it feels wrong, it probably is" - numerous Codeweavers' developers

The framework we use at Codeweavers is the excellent Moq, therefore when something is difficult to mock we are forced by the framework to write an adapter. We use an interface for testing, then create a concrete type which simply invokes the hard to test code such as static code, third party libraries and resources that are expensive to set up. There are some ways ways in C# to get around this, but they involve black magic and should be avoided at all costs unless you are deeply entangled in legacy code. A refactoring would be preferable over hard to test code.

The process of writing an adapter around hard to test code is a standard practice, we do it all the time as we are forced to by the unit testing framework. Some frameworks we use at Codeweavers such as ASP.NET MVC are designed with testability in mind, so unlike scenarios where you cannot test code easily, the MVC framework makes it possible. In a recent feature myself and a fellow colleague wrote some code within a controller which relied on some of the controllers' (the MVC framework) internals.

Rather than abstracting this into a class which we inject to make testing easier we went the route of setting up a complex, messy and tedious routing test fixture. Why you ask? Maybe it was the fact it was possible to test. Had it been straight up impossible or much harder, then introducing an abstraction would have been the obvious solution. The code in question was a small method that depending on the somewhat complex and unique routing values performed on a certain response. Fast forward a week later and the feature is to be expanded.

We were back were we started, the new feature needed more setup that relied on the framework, and in turn once this production code was changed, the old test fixture would need updating. The very thought of this made me feel tired, fed up and generally annoyed that the test code was harder to write than the actual production code! While the code did not feel right, the actual process was a by the book approach, so it must have been right. Taking a step back myself and my new pairing partner decided for a different approach. Lets abstract the controller internals we need and inject this into the controller. In turn our code would read better and the tests would be easy to construct.

Having made this refactoring the tests were still green. The refactoring was a great success. Now the test fixture set up consisted of a few simple lines. All the complex framework specific nonsense had disappeared. Getting to this stage took a bit of thought with regards the implementation, but we got there non the less. Having made this change, we wrote the next tests with such ease and joy it actually felt fun, enjoyable and completely stress free. Just how programming should be.

For the production code, as the framework is test friendly we had some unit tests around the concrete object used in production. For scenarios where this is not possible, a high level acceptance test to ensure things are wired up correctly would suffice. Either way we should always be confident when using code we do not own that it is correct, providing we use it correctly. After all, this will be heavily tested by the third party or so we hope. Manual testing will catch any integration issues with third party code with any luck.

The whole process was staggering, I was blown away by my ignorance. I knew the best practices, yet I chose to depend on concrete implementations rather than abstractions. After this session the whole theory behind mocking roles and not types [pdf] became so much clearer. This is yet one more revelation to add to the list. Every time I write Mock<name>, stop and think. Do I own the type? If not then maybe there is an abstraction waiting to escape, after all it will save a lot of pain.

MBUnit to NUnit

Over the last few weeks we've ported our tests from MBUnit to NUnit. This was done as after a quick spike it was seen that NUnit tests run almost fifty percent quicker. For example our common projects' test time went from around 40s to around 20s on average.

This whole process was no easy task. Initially our largest project was converted by the whole team. We split into pairs/individuals and tackled a test project each. Working in this manner we could commit after each project, meaning at any one time the build was only fractionally broke, rather than completely unbuildable. Previously we tried a big bang approach but after several thousand errors, we quickly reverted. After each commit the tests were gradually moved over. This took around an hour or so, and therefore our allocated dojo/technical dojo time for that week was used. For the remaining projects an ad-hoc approach was taken. The first pairs to work on a project would be responsible for porting the tests over. Thankfull our other projects bar one were fairly straightforward to upgrade and were done as part of waste or kaizen.

Some of this process could be automated however things were not completely smooth. For example converting the MBUnit namespace over was achieved by project level find and replace. Other issues such as Asserts being slightly different required a manual fix. One example being asserting a exception is thrown. The MBUnit approach used attributes while in NUnit it is more preferable to use Assert.Throws. The other issue we faced was porting over the relevant build scripts and Cruise Control configs. Again there was no easy way to do this. We had a fair few CI fails when this was done, but when editing the xml build files there is no real way to test what you've done without actual trying it!

Overall the whole episode was not as bad as I thought it would be. We seem pretty stable at time of writing, and the tests are definitely quicker to run locally. We still have slow tests, and as part of waste we'll be looking into whether these slow tests are needed. One interesting practice I've noticed over the upgrade is how many dodgy tests we've removed. Tests such as Assert.IsNotNull after creating a new object - the sort of tests everyone writes when starting TDD have been removed. These legacy tests serve no purpose now, but were the key starting point of the TDD introduction to Codeweavers several years ago. Other tests which are covered else where or simply not needed were also removed. The final issue we are aiming to improve is that of our regression/acceptance tests, many of which are Selenium tests.

Would we recommend upgrading your test suite to the latest/next best thing? Not unless you can prove with figures that it has an actual benefit. We provided no value to the business by doing this, but by hopefully taking one step to increase our feedback cycle we'll see the benefit over time. If anything, we should be more likely to run our tests. As for why MBUnit was slower? It features a lot of stuff we simply don't need, while NUnit is more lightweight and just plain faster for our use. We could perhaps speed the tests even more by writing our own test runner, but the likes of Visual Studio integration are a must therefore this is no easy task.

One interesting point to conclude was that during this process there was talk about wrapping NUnit within a Codeweavers test framework, essentially meaning we could switch test frameworks whenever. Is this overkill for most projects? Most likely, but it was something to consider especially for large applications. As who knows, maybe there will be an even faster framework out there that we can upgrade to again, next year...