Date

All I wanted was to get subunit output from our test runner. I got it, but at a much higher price than I expected to pay. This is my story.

Quick and dirty, April 2009

It’s often a good idea to try to get something working as quickly as possible, even if the hack to get it working is a little dirty.

For getting subunit output from our test runner, the right thing to do is to change the TestResult object used by the test runner. In Python’s standard unittest and in most xUnit frameworks, TestResult objects are responsible for reporting results to the user.

If zope.testing were designed more in-line with other Python testing frameworks, this would also have been the quick and easy way to do it.

Sadly, zope.testing embeds quite a lot of logic necessary for the execution of tests in its TestResult object, so I can’t just swap it out for one that implements the standard interface.

Instead, I created a MultiTestResult object and added it to testtools. It acts like a single test result but dispatches to a bunch of test results.

Then I did a quick and dirty thing to silence any output that zope.testing’s result object might have generated, and used MultiTestResult to glue in a subunit test result. Since zope.testing doesn’t have a way of plugging in test results, I did some evil monkey patching.

Fix zope.testing, 5th February 2010

Now, what I should have done is fix zope.testing so that it had its own way of generating subunit output cleanly.

I didn’t, because at the time Launchpad was using a very old version of zope.testing. We couldn’t upgrade because we were using Python 2.4 and new Zope bits needed Python 2.5.

However, eventually we got Python 2.5 support, upgraded our Zope Toolkit bits and pieces and started using a newer version of zope.testing. Once that happened, patching zope.testing became feasible, since I’d actually be able to use the results of my labor.

Looking at zope.testing, I would have to either disentangle the layer-running logic from the TestResult code, or I would have to implement my own “formatter”. Disentangling the layer stuff would be way, way too much work, so I went for the formatter.

A zope.testing OutputFormatter is an object that implements a very big, undocumented interface and is responsible for displaying almost all of the output that the Zope testrunner could possibly generate.

To implement one, you have to implement twenty-seven methods, most of which lack clear documentation or usage information. There are tests, which is nice, but the tests are big integration tests, rather than unit tests.

Anyway, I implemented a subunit output formatter over the course of a few weeks, then submitted it to the Zope community and got it landed. Really, they are wonderful people.

Next step, actually using this patch.

Upgrade zope.testing, 13th March 2010

Upgrading zope.testing is easy enough. We use buildout, it’s a simple question of updating a version number in a configuration file, running a few simple commands, wondering why it doesn’t work, discovering it’s a local config issue, repeating two or three times and then you’re done.

Upgrading subunit was really hard. To explain why, I’ll have to go into some detail about the way Launchpad handles dependencies.

We have three different kinds of dependencies:

  1. Source code dependencies
  2. Buildout dependencies
  3. Package dependencies

Each of them is managed differently, and updating each of them has its own complexities. I ended up trying all of them.

Source code

Subunit was already included as a source code dependency. That means that we had a branch with the Launchpad version of subunit, and that we pull from that branch whenever we roll out or update our code. Kind of like svn:externals, but managed with custom scripts.

I needed to use a newer subunit than the one we already had in order to get zope.testing subunit output working. To do that, I would have to merge in changes from subunit trunk. Normally, that would be quite easy.

However, subunit has since upgraded its Bazaar repository format from something old and crappy into the new, shiny and awesome 2a format. The 2a format is completely incompatible with older formats, and I cannot merge from the new trunk into Launchpad’s old subunit branch.

I have no idea still how to upgrade Launchpad’s old subunit branch. It’s a PQM-managed branch and I couldn’t find documentation. Even if I could upgrade the branch, I’d have no promise that the tools that our ops guys use to rollout Launchpad would be robust in the face of a Bazaar repository format change. The last thing I want to do is break a rollout because I want to upgrade some developer tools.

Buildout

I then gave up on using source code dependencies and tried buildout, which is our “recommended” way of handling dependencies.

subunit is a tricky thing to handle with buildout. Buildout is at its best when it’s managing Python packages. subunit is not a Python package. It’s a multi-language project that builds into some binaries and some libraries, including Python libraries. It uses autotools to do all of this.

There are ways to build autotools packages using buildout. You add the egg for the “cmmi” recipe to your dependencies and specify a particular build recipe for the autotools package. This kind of worked, but it left me with two problems.

The first was that the Python libraries that it generated were buried somewhere deep in the build output, and not included naturally in the Python import paths as happens with a regular setup.py build. Not a big deal, we can glue it together with symlinks.

The second is that to build subunit, I actually needed cppunit and check, and maybe some other things. I really, really don’t want to manually traverse the build dependency chain of subunit and add each of these things as eggs to our buildout just so I can get subunit output from our test suite.

Packages

Which naturally made me think of Debian packages. After all, what better way to manage complex dependencies?

Unfortunately, I really don’t have any idea how to update the Debian packages in our PPA. I asked on #launchpad-dev and quickly got some helpful advice about how to do it.

There’s some debian/ directory fiddling, changelog updating and so forth. Then I use bzr-builder. Then I submit the branch to lp:meta-lp-deps, which is where we manage the code for the package, then I build it in the PPA. The PPA build doesn’t work because I didn’t sign the code of conduct – easily fixed.

I try again and it works and it looks like all is right with the world. Apparently though, I got something wrong. Luckily maxb fixed it for me while I slept so I didn’t need to do anything about it.

Yay! I have now officially updated the version of subunit that we depend on. Now all that’s left for me to do is propagate that change and then land my branch.

To propagate the change, I need to update the EC2 images that we use for running our tests. There’s a command to do that and a very helpful wiki page with instructions.

I follow the instructions, and it doesn’t work. The error is weird, and to do with some crazy socket thing. No one has a clue what to do about it, and I’m actually quite busy doing other things, so I leave it rest for a couple of days.

When I come back to try again, I google around for the error and discover that it’s actually a bug in openjdk that affected the Lucid beta release and has now been fixed. Hooray, I guess.

I update my Lucid install and get ec2 to build the new image for testing. Once that’s done, I need to get our ops guys to update the completely different set of images that we use on our buildbot.

I send off a request, and there’s a bit of back-and-forth. Apparently we use one archive for development, which I’ve updated, and a completely different archive for production rollouts, which I am not allowed to update. After ten working days, it all gets sorted out, and now I am able to land my branch.

Landing the branch, 10th April 2010

At last, I’m ready to land my branch.

It doesn’t work. Of course.

I could have sworn that I ran the full test suite with it, both with subunit output and with the default, but it seems to be broken in two ways.

The first is that the subunit output support in zope.testing is broken. I made the error method raise NotImplementedError. This means that if ever a layer fails to start up properly, the test runner dies with an unhelpful error that masks the layer’s own error. Suck.

The right way to fix it is to add a new API to the formatter to specifically handle layer set-up errors. I take the quick-and-dirty approach of just printing out whatever error gets. Patch is sent off to zope-dev mailing list.

The second is that new zope.testing has deprecated its doctest module, and emits deprecation warnings all over the place. We have tests that fail if warnings are emitted – sixteen, to be precise – so I can’t upgrade until I somehow stop the warnings.

They are generated by zope.app.testing.functional in the Librarian start-up. Why on earth the Librarian needs zope.app.testing is beyond me. I fetch zope.app.testing, work around a bug in python-setuptools to get it building, patch it and submit the patch to the zope list.

If I wanted to, I could make two new eggs now for the two patched projects and then just land the branch. I don’t think I want to, since the patches are so small I’m confident I can get them landed and maybe even persuade someone to do releases.

The fix to zope.app.testing lands without a hitch, and yet another kind person from the Zope community does a release and even do what they can to get the other patch landed.

However, it turns out that between zope.app.testing 3.7.3 and the new 3.7.5 that I need, zope.app.testing has started to depend on [zope.component](http://pypi.python.org/pypi/zope.component) 3.8 or better. zope.component has made some changes that break API compatibility with a stack of other Zope bits, including [zope.sendmail](http://pypi.python.org/pypi/zope.sendmail) and [zope.app.component](http://pypi.python.org/pypi/zope.app.component). zope.sendmail has been updated, but zope.app.component has not. I can’t find the code to patch it either.
Luckily, Launchpad doesn’t really use the broken bits of zope.app.component, so it’s easy enough to [work around the problem](http://bazaar.launchpad.net/~jml/launchpad/new-zope-testing/revision/10456) there, and bump zope.sendmail’s version in our buildout configuration.
Then the fix for zope.testing lands, and [Sidnei](https://launchpad.net/~sidnei) is again kind enough to release. I bump our versions.cfg, and land the branch.
It is finished.
Thoughts ———— It should be obvious by now that I think the whole process has been unreasonably complicated and taken far too long. Perhaps I chose a task that was fundamentally difficult. Perhaps it was a one-off “perfect storm” of confounding factors. Perhaps there are deep, serious problems in the way we do things. Perhaps I’m just not very good. I certainly made some mistakes. For a start, I shouldn’t have been so afraid of breaking the edge rollout. We take measures like daily edge rollouts precisely so that we can catch our mistakes, which should embolden us to try things. I also should have done more testing between the time I asked for the image updates and the time I landed the branch.
I kind of like the way that Zope has been split out into a million little projects. I think it made it easier for me to come along with my small, opportunistic patches, and made it easier for others to review and land them. It also makes it easier for stupid backwards compatibility mistakes (since you can’t just grep a tree). I somehow think that if [Twisted](http://twistedmatrix.com/) broke itself up, they’d find a way to deal systemically with backwards compatibility changes, if only by blocking them in review.
Also, it would be pretty much impossible to use Zope at all without [PyPI](http://pypi.python.org/).
I would love to know what you all think of this, and whether you have any thoughts on how we can make hacking on Launchpad easier.