An introduction to the Debian Continuous Integration project

June 01, 2014
Tags: ,

Debian is a big system. At the time of writing, looking at my local package list caches tells me that the unstable suite contains 21306 source packages, and 42867 binary packages on amd64. Between these 42867 binary packages, there is an unthinkable number of inter-package dependencies. For example the dependency graph of the ruby packages contains other 20-something packages.

A new version of any of these packages can potentially break some functionality in the ruby package.

And that dependency graph is very small. Looking at the dependency graph for, say, the rails package will make your eyes bleed. I tried it here, and GraphViz needed a PNG image with 7653×10003 pixels to draw it. It ain’t pretty. Installing rails on a clean Debian system will pull in another 109 packages as part of the dependency chain. Again, as new versions of those packages are uploaded the archive, there is a probability that a backwards-incompatible change, or even a bug fix which was being worked around, might make some funcionality in rails stop working. Even if that probability is low for each package in the dependency chain, with enough packages the probability of any of them causing problems for rails is quite high.

And still the rails dependency chain is not that big. libreoffice will pull in another 264 packages. gnome will pull in 1311 dependencies, and kde-full 1320 (!).

With a system this big, problems will arrive, and that’s a fact of life. As developers, what we can do is try to spot these problems as early as possible, and fixing them in time to make a solid release with the high quality Debian is known for.

While automated testing is not the proverbial Silver Bullet of Software Engineering, it is an effective way of finding regressions.

Back in 2006, Ian Jackson started the development of autopkgtest as a tool to test Debian packages in their installed form (as opposed to testing packages using their source tree).

In 2011, the autopkgtest test suite format was proposed as a standard for the Debian project, in what we now know as the DEP-8 specification. Since then, some maintainers such as myself started experimenting with DEP-8 tests in their packages. There was an expectation in the air that someday, someone would run those tests for the entire archive, and that would be a precious source of QA information.

Durign the holiday break last year, I decided to give it a shot. I initially called the codebase dep8. Later I renamed it to debci, since it could potentially also run other other types of test suites in the future. Since early January, ci.debian.net run an instance of debci for the Debian Project.

The Debian continuous Integration will trigger tests at most 4 times a day, 3 hours after each dinstall run. It will update a local APT cache and look for packages that declare a DEP-8 test suite. Each package with a test suite will then have its test suite executed if there was any change in its dependency chain since the last test run. Existing test results are published at ci.debian.net every hour, and at the end of each batch a “global status” is updated.

Maintainers can subscribe to a per package Atom feed to keep up with their package test results. People interested in the overall status can subscribe to a global Atom feed of events.

Since the introduction of Debian CI in mid-January 2014, we have seen an amazing increase in the number of packages with test suites. We had little less than 200 packages with test suites back then, against around 350 now (early June 2014). The ratio of packages passing passing their test suite has also improved a lot, going from less than 50% to more than 75%.

There is documentation available, including a FAQ for package maintainers with further information about how the system works, how to declare test suites in their packages and how to reproduce test runs locally. Also available is development information about debci itself, to those inclined to help improve the system.

This is just the beginning. debci is under a good rate of development, and you can expect to see a constant flux of improvements. In special, I would like to mention a few people who are giving amazing contributions to the project:

  • Martin Pitt has been working on improving debci to support parallel and distributed workers. Being the current autopkgtest maintainer, Martin also already got some bug fixes and fixes into autopkgtest motivated by debci use cases.

  • Brandon Fairchild is a GSOC student working on improving the debci web interface to provide more useful information, display information for multiple suites and architectures, plus making the UI work even without Javascript enabled.

  • Lucas Kanashiro is another GSOC student, who is working on investigating patterns among packages that fail their test suites, so that we can figure out how we can fix them, or if there are classes of failures that are actually caused by problems in the debci infrastructure.