Upcoming CI changes - transition to VM based CI

[Albert Astals Cid]

El dimarts, 3 de juny del 2025, a les 11:42:08 (Hora d’estiu d’Europa 
central), Ben Cooksley va escriure:
> On Tue, Jun 3, 2025 at 9:03 AM Albert Astals Cid <aacid at kde.org> wrote:
> > El dilluns, 2 de juny del 2025, a les 13:39:21 (Hora d’estiu d’Europa
> > 
> > central), Ben Cooksley va escriure:
> > > Hi all,
> > > 
> > > For some time now we have had a variety of issues with our Docker/Podman
> > > based CI builds. These have included the lack of GUI test support on
> > > Windows, periodic crashes on FreeBSD, poor IO performance of Windows
> > > builds, issues supporting builds for Flatpak and Snaps and inability to
> > > support either builds or tests where elevated privileges or system
> > 
> > session
> > 
> > > resources are needed.
> > > 
> > > In addition to this we've had issues where Linux CI builds have the
> > > capability to trigger OOM events on the CI hosts, which in turn takes
> > > out
> > > Windows and (less often) FreeBSD builders. While this does not occur too
> > > often, it does happen from time to time and eventually negatively
> > > impacts
> > > the build queue for those platforms.
> > > 
> > > The need to have dedicated VMs for FreeBSD and Windows on our builders
> > 
> > also
> > 
> > > makes setting up of a CI build node for KDE software a more complicated
> > 
> > and
> > 
> > > time intensive task than it otherwise needs to be (and means that the
> > > amount of systems to care for increases by 3 for every CI node we add).
> > > 
> > > While individually relatively minor, together these issues more than
> > > justify making a significant change to the way we run our CI system - in
> > > this case transitioning from container based builds to VM based builds.
> > > 
> > > These builds will still take place on dedicated hardware that we
> > > control,
> > > however instead of taking place within a container (managed by Podman on
> > > Linux and FreeBSD, or Docker on Windows) they will instead take place
> > > within a VM using a copy-on-write disk image.
> > > VM based builds will unfortunately take a little longer to start (it
> > 
> > takes
> > 
> > > ~10 seconds for a VM from any of Linux, FreeBSD or Windows to boot on my
> > > personal system) however the benefits we gain should more than outweigh
> > > this small downside.
> > > 
> > > This has been under development for the past couple of weeks and is now
> > > reaching the point where the only remaining steps are to get it
> > 
> > integrated
> > 
> > > with the Gitlab CI agent (gitlab-runner) for which prototype code is
> > > already in place, and complete porting of our images over. Once that
> > > happens a complete rebuild of all of our builders will be swiftly
> > > undertaken to transition them completely over to the new VM based
> > > infrastructure.
> > > 
> > > Specs wise, at this time it is planned for each spawned standard VM to
> > > be
> > > provided with 2/3's of the system CPU cores (so 12 cores), 16GB RAM and
> > > 100GB of disk space (although some of that will be occupied by the
> > > system
> > > image - approximately 10GB for standard Linux builds and ~30GB or so for
> > > Windows builds). There will be a higher resource tier available for
> > 
> > certain
> > 
> > > builds however that will be on request only and would need to be
> > 
> > justified
> > 
> > > (such as Craft needing to build QtWebEngine).
> > > 
> > > As launching VMs is not the most efficient approach for all workloads,
> > > limited support for running Docker containers will be preserved, however
> > > this support is primarily intended for running linters, sanity checks
> > > and
> > > website builds, and is not intended for running general CI/CD builds.
> > > 
> > > The tooling used by the CI nodes to run VMs is something that should be
> > > fairly trivial for people to run on their own local system should they
> > 
> > wish
> > 
> > > to run any of those images (say for FreeBSD or Android), although you
> > 
> > will
> > 
> > > need to setup libvirt yourself (SUSE has very good instructions for
> > > this,
> > > Debian less so as their instructions lack installing the packages needed
> > 
> > to
> > 
> > > provide UEFI and TPM support). The tooling itself was merged this
> > > evening
> > > to sysadmin/ci-images (vm-common/ folder) and can be used with the VM
> > > images found at https://storage.kde.org/vm-images/
> > > 
> > > There is however one downside to this - Qt 5 support.
> > > 
> > > Over the past few months distributions have been steadily removing
> > 
> > packages
> > 
> > > and other supporting infrastructure needed to keep Qt 5 builds alive. In
> > > the case of Windows, support for the entire Qt 5 tree has been
> > 
> > unmaintained
> > 
> > > for some time. For FreeBSD and SUSE a significant number of packages
> > > have
> > > been removed - which in the case of SUSE also includes packages needed
> > > to
> > > support the building of KJS.  Accordingly, because builds of Frameworks
> > 
> > are
> > 
> > > a first stepping stone to support building anything else, it will not be
> > > possible for us to produce Qt 5 based VM build images for any of the 3
> > > platforms.
> > > 
> > > We will therefore have to remove Qt 5 support from the CI system with
> > > the
> > > transition to VM based CI.
> > 
> > From previous discussions I had the impression this was only for things
> > that
> > wanted to create packages and not for "want to have CI to compile/run
> > tests".
> > 
> > Can you confirm you are proposng a total annihilation of Qt5 support in
> > our
> > CI?
> 
> At the time we had that discussion it was still possible to build some of
> the Qt 5 images, however that is no longer the case - all of them now fail
> to build.
> 
> In the case of the suse-qt515 image, the removal of libpcre in SUSE means
> it is no longer possible to build KJS.
> Consequently, we're no longer able to build all Frameworks (making 'kf5'
> branch CI for Frameworks non-functional) so there isn't much point in
> looking further to support Qt 5 on Linux.
> 
> For FreeBSD the story is much the same as SUSE - packages are being removed
> as apps upgrade to Qt 6 and the Qt 5 version of libraries becomes surplus
> to requirements.
> 
> For Windows the continued operation of that CI has only been possible
> because our existing images are still around - new ones cannot be built.
> That has been the case for a significant amount of time now, and it is not
> worth the investment to fix it as everyone who works on Windows has moved
> on to Qt 6.
> 
> In essence there is little we can do to keep this alive - distributions are
> removing support so we must follow suit.
> 
> The correct course of action is to accelerate the porting of the remaining
> applications, not to delay and keep Qt 5 alive.

As Nico said, we can just not build KJS/KHTML and then give the projects that 
still use Qt5 a reasonable timeframe for them to port to Qt6, e.g. until the 
end of the year, not until the end of the month.

Cheers,
  Albert

> 
> > Cheers,
> > 
> >   Albert
> 
> Regards,
> Ben
> 
> > > Please let me know if there are any questions on the above.
> > > 
> > > Thanks,
> > > Ben