Copr Modularity in retrospect

12. 01. 2018 | Jakub Kadlčík | EN copr fedora modularity

This article is about the journey that we made since the Fedora modularity project started and we decided to get involved and provide modularity features in Copr. It has been a long and difficult road and we are still not on its end because the whole modularity project is a living organism that is still evolving and changing. Though, we are happy to be part of it.

First demo

More than a year ago, in let’s say the dark ages, nothing existed. Just the void, … well ok, enough metaphors. At that time, the whole modularity was just an idea, any real implementation existed. Then we came with possibly one of the first prototypes. It was a simple feature that allowed a user to assign a modulemd yaml file to a chroot in the project settings. When packages were built in such project/chroot and the repodata was created, we added the modulemd to them. This was the first piece of modularity code in Copr and the beginning of our journey.

The “webform”

A decision needed to be made, which audience we want to target to? We realized that the modularity team (later known as factory 2.0) will focus at first on experienced users, i.e. themselves. For us, it meant that the whole space targeted on complete beginners will not be covered soon, so we decided to focus on this area.

Ask yourself a question. If your entire community is people with diverse, but at least fundamental knowledge of RPM packaging, what is the most challenging task that they need to do, to be able to distribute their software as a module? They all need to learn how to write a modulemd yaml file. It isn’t a rocket science, but it is time-consuming and also at that time the only guide was a documented template file. This could be enough to discourage a lot of people. We saw a chance to attract people by simplifying things.

The idea of “the webform” was born. We implemented a form for creating modules without any knowledge requirement. A user could just specify the module name, version, and release (currently name, stream, and version) and select from all the packages built in the particular project, which should be part of the module. Then Copr parsed it, constructed a yaml via the python-modulemd library and it continued the pipeline so the module was built (accordingly to what was considered “to be built” at that time). This was a killer feature for Copr. You can see a demo here.

DNF 1.x

Copr was able to build modules and generate repofiles for them. However, nobody could be sure, whether it was done correctly and if the output was in expected format, because there was no official way, how to install modules on the client machine. To be precise, some prototypes existed. Modularity features for DNF were implemented as the fm-dnf-plugin and fm-modulemd-resolver. There was a little problem though - it didn’t work with the most recent version of the modulemd format. We decided to update the code in order to use the DNF plugin in our tests to ensure that we produce the module repository in a way that we are supposed to. The funny thing is that meanwhile the DNF plugin was stated as obsolete in favor of DNF 2.0 with internal support for modularity. Hence, my pull requests PR#5 and PR#1 were destined to be forever unmerged.

Module Build Service

aka MBS, aka fm-orchestrator, formerly known as Říďa (from Czech comic book Čtyřlístek), is a service for orchestrating a module build in configured buildsystem, such as Koji, Copr or Mock.

As the complexity of building modules grew it was a particularly good idea not to re-invent a wheel and use/improve tools that Factory 2.0 already created. The most important for us was the MBS as it should have been able to resolve module build dependencies, configure the buildroot and schedule the builds in the right order. Moreover, the logic was abstracted to general code that was common for all builders, so from our perspective it was a handy little blackbox managed and internally reworked by people who defined the Modularity.

It was on us to implement the Copr builder and deploy our own instance of this service. As we quickly found out, we were the first ones who wanted to do so. There was no ansible playbook for deployment and moreover, it wasn’t even packaged for Fedora. Packaging was eventually done by the upstream and we came up with the playbook. You can see the demo of our first builds via MBS here.

Buildroot from modules

As you can see from the previous demo, we used to build modules in Fedora rawhide chroot. This contradicted with one of the former modularity goals which was a fully modular system built in fully modular buildroot. We solved it by building modules in custom-1-x86_64 chroot with an enabled external repository to base-runtime (later platform) module. You can see the demo here.

We took it even further. Although the MBS constructs a buildroot from modules, it actually kind of cheats. It determines which modules should be installed into the buildroot and then finds out which packages are provided by those modules. Those are installed into the buildroot. This makes a sense maybe as a workaround, but I consider it wrong from the design perspective. This made us implement a support for Mock to be able to directly install modules into the buildroot and then updated MBS so it specifies dependencies to the Copr as modules instead of packages.

Modules with Copr packages

How to built modules from Copr packages? Wait for an upcoming article. It will be very short with a lot of images :-).

MBS Plugins

It turned out that having the code of third-party builders such as Copr in the fm-orchestrator repository is quite ineffective. The Copr release process became dependant on MBS release and at the same time, the MBS upstream didn’t know the implementation details of the Copr side. We mutually agreed that custom builders should be moved out to their own repositories in the form of plugins. That day module-build-service-copr was born.

Hybrid Modularity

Just before Christmas break a highly awaited article Modularity is Dead, Long Live Modularity! was published. If you haven’t read it yet, I recommend you to do so. It explains how the Modularity design is going to be changed in the foreseeable future. From the Copr perspective, the most important revelation is that the minimal buildroot is not going to be composed of modules anymore. Instead, the standard Fedora buildroot will be used.

Conclusion

What to take from this already too long article? We very much appreciate your feedback. Things regarding modularity change very fast and very often and it is sometimes difficult for us to find the right way how to do things and to catch up the hot news. Your opinions and ideas help us to prioritize the agenda, speed up the process of figuring out what are the best solutions for us and ultimately determine the direction which we should take in the future. Thank you for that.

In the following article, we are going to talk about how you can currently build modules in Copr and how the current user interface looks like.