Nicolas Hicher is the XiVO release manager that make it happen. His input comes from a number of sources which are gradually migrated toward a set of git repositories which will become a central place:
- git://git.xivo.fr/xivo-dalek.git for old stable
- git://git.xivo.fr/xivo-gallifrey.git for stable
- git://git.xivo.fr/xivo-skaro.git for unstable
- http://git.proformatique.com/ debian fai and XiVO Qt, Web, Android clients and more
About ten developers have write access to commit their work on these repositories. In addition to their daily workload, they integrate patches and bug fixes contributed to the XiVO redmine. Nicolas maintains packaging tools deployed on a few virtual machines which purpose are to produce a useable Debian repository. Although it is currently updated with dak it is complicated to maintain and will be migrated to reprepro which is easier to deploy and maintain.
The general idea is to follow the Debian philosophy with regard to maintaining stable distributions :
- the unstable (skaro) distribution roughly matches the Debian testing distribution: it has been tried to be installable and roughly functional, but it contains newly implemented features that may require bug fixing when facing real world usage.
- the stable (Gallifrey) distribution is only updated for security purposes, with the exception of newer asterisk versions.
- the old stable (dalek) distribution is kept for historical purposes
The Debian unstable equivalent is the playground of the ten developers and there is no public URL for it. When one of them wants to update a package that is being actively developed, he logs in one of the virtual machines maintained by Nicolas, types a few command to create "dev" packages. All the developers use deb packaged based operating systems (Debian GNU/Linux or Ubuntu) and their source list points to the repositories located on the virtual machines where the "dev" packages are created. When one of the developers breaks a package, the others are quick to pick on it. This is both a curse (when it disrupts their work) and a blessing (because there is a continuous check of the distribution consistency). In order to help maintain the consistency of the distribution, strict version dependency is enforced between packages : if version 1.4.6 is required, version 1.4.7 will be rejected although it is presumably backward compatible. This policy may be relaxed in the future at the cost of extra care to ensure backward compatibility.
When the developers are satisfied with the state of the XiVO distribution, they declare it a release and set a corresponding git branch (formerly a SVN tag). Nicolas notifies the production (i.e. the people ensuring that all deployed XiVO installation can rely on the stability of the existing repositories) that the skaro distribution (i.e. http://dak.proformatique.com/debian/dists/lenny-xivo-skaro-dev/) will be broken while the development packages are migrated to skaro. He then proceeds with the update and tests the resulting repository by performing an installation from scratch on a virtual machine via PXE. If the virtual machine indeed contains the expected packages and that the web interface displays, the production team is notified and does additional tests before launching upgrades where relevant.
When a packaging problem is found, either by the production team or by the developers, Nicolas is responsible for fixing it. If a new software is being deployed, Nicolas is also responsible for creating the corresponding package from scratch. More often than not, a problem requires an emergency fix by the developers shortly after a new release is put in production. In which case the whole process outlined above needs to be done again.
The packaging process is evolving continuously and the following improvements are being considered:
- adding RPM based packages
- adding support for x86_64 and leave room for more architectures
- deploy a reprepro repository to replace http://dak.proformatique.com/debian
- publish and document XiVO packaging tools
- Debian GNU/Linux squeeze support
- move as much as possible from the package to the source tree : some packages contain so much knowledge that installing from sources is difficult and is a serious obstacle for people running distributions which are not explicitly supported
- improve and document the creation of CD based distributions and ready to use virtual machines
- if a single machine is used to build packages, allow to build multiple packages in parallel
With these in mind, the following use case should be available to XiVO developers and packagers:
- developer side: I'm working alone on my code, it is perfect, ready for production. I build the package (or the whole corresponding suite) and I publish the package in the corresponding development repository, so that it is available for other developers.
- developer side: I'm not the only one working on the code (or I know my work is not satisfactory) but I want to test a killer feature. I build the package and publish it on my private repository and I can test it without impacting other developers.
- packager side: the developers did a lot of work and there are many bug fixes, asterisk version 3 has been published and chan_sccp is stable. I build a full release ( tag + build + publish on a release candidate repository ). The production team runs tests suites and find bugs that are patched by the developers and propose new packages. When the loop stabilizes, the release candidate replaces the production.
- packager side: a new package is to be built, it must be added on all repositories. The developer responsible for the code provides all the information to:
** build the package (dependencies, location of all files etc).
** a HOWTO explaining how to test that the resulting software is working. It needs not be 100% tests but should cover the basics. A command --version returning the version number is better than nothing.
The analysis of the existing XiVO release management practices and prospects reveal patterns that are common to a number of software publishers. At least those who are best served by deploying a dedicated distribution (repository, ISO or VM) rather than a single software package. They are in many ways identical to wide spread distributions such as Ubuntu or Debian GNU/Linux with a fundamental difference: it is cross distributions (Fedora, Gentoo, Debian GNU/Linux, Suse ...).
The packaging-farm tool has been developed to address this specific issue and will be used as a base to help implement the above features. It is based on the following ideas:
The input of the farm is a package and its output are repositories for the target distributions.
A Debian GNU/Linux package is a well defined format to transport software and packaging instructions. Using it to feed the farm saves the need to define custom conventions. The same goes for the output of the farm : all distributions define a format to store collections of packages, in source and binary forms.
The environment of a failed package attempt is preserved for debugging.
Each package is built in an pristine environment to minimize side effects due to previous executions. This environment (chroot or virtual machine) is destroyed if the packaging operation is a success. It is preserved if the package fail and the maintainer can chroot or ssh into it to figure out and fix the problem.
A farm can be stacked on top of others.
This really is a property of the fact that each farm creates a well formed repository for each distribution. A given farm can either build a package that has been submitted to it and put the result in its output repository. Or it can copy the content of the repository of a child farm into its own. For instance, a production farm/repository can be made a parent of a release candidate farm/repository. When the release candidate repository is declared final, the production farm/repository copies the content of the release candidate farm/repository into its own.
A farm delegates the work to a slave machine.
The farm creates environments (aufs based chroots and possibly lxc or kvm based virtual machines) and submits a source package for building. It does not assume the native operating system is able to build any kind of package itself. If a farm runs on a given operating system, it may be easier to delegate the actual building of the package to a virtual machine running the native operating system for the targeted distribution.
All package specifications must be maintained manually.
It is difficult to create a package from a given distribution automatically based on the package specification from another distribution. The farm does not support this kind of tool.
The implementation is heavily based on Makefiles because it mostly handles a chain of dependencies for which they are best suited. The supporting scripts are shell based because it is the language of choice for Makefiles.
The XiVO client is being packaged by kaou (nick on irc.freenode.net) and is not covered here. Nor is the process to create virtual machines or ISO based distributions.