Test process improvements

Yesterday and today I’ve been working on improving the automated testing process used when building products like Synergy Advance and Hextrapolate.

Goodbye failing tests

I had some failing tests in place (WO_TEST_FAIL) to serve as reminders to go ahead and implement real tests at a later date. This is mostly in cases where the tests are fiendishly difficult to write (testing GUI elements, for example) and I wasn’t sure how to proceed.

So I’ve gotten rid of these placeholder tests and put good old-fashioned TODO comments in their place. This means that I can now do builds and expect 100% success rates; failure no longer means "look into this later", it now means "here’s a problem that needs to be fixed right now". This is a change that I probably should have made a long time ago.

This in turn makes my automated nightly builds process more robust. Instead of manually reviewing the builds before allowing the nightly to proceed I am one step closer to letting them run truly automatically (for the time being I am still firing them off manually until I am 100% convinced that the process is bullet-proof).

Notifications

The next thing I want to do is show Growl notifications when test runs finish, using the growlnotify CLI tool. In the case of nightlies I’ll make these "sticky" notifications so that they’ll hang around on the screen even if I’m not in front of the computer at the time the nightlies run. In the case of normal builds I’ll make them non-sticky.

Truly continuous integration

My next step will then be to set up truly continuous automated testing for better continuous integration. The idea is to perform the following in the background in an endless loop:

  • Checkout an up-to-the-second copy of the project source code.
  • Perform a full build at a lower priority (using nice).
  • If any errors occur or unit tests fail, notify using growlnotify.

The idea is that instead of thinking about doing a change-build-test cycle you just think about changes: change, change, change, change, continuously advancing and letting the background tests notify you if one of your changes needs revision.

Thanks to SVK (which provides a local mirror of the remote Subversion repository) this cycle can be performed without hitting the network. Thanks to nice all this should be possible without bogging down the machine (although more memory would be nice, I admit).

It shouldn’t be too much work to set this up; I can probably just modify the current nightlies script and pass in a command-line switch to get the desired behaviour.

This is just one of the ways in which my recent work using Ruby to write Walrus has made me a better programmer; working with the excellent Autotest suite prompted me to take this step.

Packaging tests

Finally, I want to add some more tests that focus on the built product (testing the build process) rather than on the code (testing the code). I’ve found issues in the past where a piece of the program didn’t work properly because of an issue external to the code; for example, an Info.plist file that was incorrectly encoded as UTF-16 rather than UTF-8 and which prevented a preference pane from being included in the list of available panes (but only in one of the localizations).

It might have been possible to catch this with normal unit tests, but only if I ran them all for each localization (which is something I should consider doing in the future anyway). This experience has happened more than once; it is all too easy for encodings to get mucked up in Xcode because of carelessness.

So I want to add some "packaging" tests to inspect the built product before it’s considered satisfactory. Things like automatically checking the encoding on plist and strings files, making sure no expected items are missing from the final bundle, making sure no items which shouldn’t be in the bundle are present (private headers, for instance), and so forth.

RubyCocoa and RSpec

Those are exactly the kinds of tests that would be easiest to write in Ruby. In fact, I think that many, many unit tests would be easier to write in Ruby. Apple is going to up their support of scripting bridges in Leopard and that means at some point in the future it is going to get a whole lot simpler to bridge between Objective-C and Ruby.

Ruby is great for writing tests for two reasons:

  1. Everything is an object: this means that you can specify expectations on literally anything at all, even Fixnum instances like 1 and 3, and special values like nil, true and false.
  2. The amazing RSpec BDD framework which leverages the everything-is-an-object design of Ruby to improve the readability of your specs:
# the RSpec way:
thing.should == "foo"

// the Objective-C way:
WO_TEST_EQ(thing, @"foo");

There’s nothing wrong with the Objective-C way (using WOTest) of course, but the RSpec way is nicer. Note that it works no matter what thing is; it can even be nil or false and the test will work as you would expect. The WOTest version will work too regardless of what thing is (it can be a NULL pointer, for instance) but it’s not as elegant.

When I saw RSpec I wanted to modify WOTest to use a similar syntax (should and shouldNot messages added via a category on NSObject) using Higher Order Messaging to allow expressions like:

[[thing should] equal:@"foo"]; But this could fall down if thing were nil and wouldn’t work if thing were not an NSObject subclass.

Note that I use a should method that returns an anonymous proxy that captures the equal: method and then forwards it back to thing rather than using a single shouldEqual: method. This is to avoid cluttering the NSObject namespace.

I guess it could be done, but you could never go the full distance in Objective-C. That’s one of the reasons I’m looking forward to Leopard; I’m hoping that the RubyCocoa bridge will allow me to write specs in Ruby that can be used to test Objective-C objects. Maybe using RSpec itself, or providing a bridge layer to tap into the existing WOTest machinery.