This repo contains a proof-of-concept tool that implements most of the requirements for KONFLUX-2110.

There are no stability guarantees.

The output should generally be compatible with what is being implemented in cachi2, but there are some differences, like adding additional sourcerpm key for binary packages to make it easier to map the package to corresponding source.

Install with pip directly from Git:

$ pip install --user https://github.com/konflux-ci/rpm-lockfile-prototype/archive/refs/heads/main.zip

Or latest released version:

$ pip install --user https://github.com/konflux-ci/rpm-lockfile-prototype/archive/refs/tags/v0.4.0.tar.gz

The tool requires on dnf libraries, which are painful to get into virtual environment. Enabling system packages makes it easier.

Additionally, the tool requires skopeo and rpm to be available on the system.

$ python -m venv venv --system-site-packages
$ . venv/bin/activate
(venv) $ python -m pip install -e .
(venv) $ rpm-lockfile-prototype --help
usage: rpm-lockfile-prototype [-h]
                              [-f CONTAINERFILE | --image IMAGE | --local-system | --bare | --rpm-ostree-treefile RPM_OSTREE_TREEFILE]
                              [--flatpak] [--debug] [--arch ARCH] [--outfile OUTFILE]
                              [--print-schema] [--allowerasing]
                              INPUT_FILE

positional arguments:
  INPUT_FILE

options:
  -h, --help            show this help message and exit
  -f CONTAINERFILE, --containerfile CONTAINERFILE
                        Load installed packages from base image specified in
                        Containerfile and make them available during dependency
                        resolution.
  --image IMAGE         Use rpmdb from the given image.
  --local-system        Resolve dependencies for current system.
  --bare                Resolve dependencies as if nothing is installed in the target
                        system.
  --rpm-ostree-treefile RPM_OSTREE_TREEFILE
  --flatpak             Determine the set of packages from the flatpak: section of
                        container.yaml.
  --debug
  --arch ARCH           Run the resolution for this architecture. Can be specified
                        multiple times.
  --outfile OUTFILE
  --print-schema        Print schema for the input file to stdout.
  --allowerasing        Allow erasing of installed packages to resolve dependencies.
(venv) $

The input file tells this tool where to look for RPMs and what packages to install. If not specified, rpms.in.yaml from current working directory will be used. It's a yaml file with following structure.

contentOrigin:
  # Define at least one source of packages, but you can have as many as you want.
  repos:
    # List of objects with repoid and baseurl
    - repoid: fedora
      baseurl: https://kojipkgs.fedoraproject.org/compose/rawhide/latest-Fedora-Rawhide/compose/Everything/$basearch/os/
      # You can list any option that would be in .repo file here too.
      # For example sslverify, proxy or excludepkgs might be of interest
  repofiles:
    # Either local path or url pointing to .repo file
    - ./c9s.repo
    - https://example.com/updates.repo
  composes:
    # If your environment uses Compose Tracking Service (https://pagure.io/cts/)
    # and you define environment variable CTS_URL, you can look up repos from
    # composes either by compose ID or by finding latest compose matching some
    # filters. Fedora doesn't use CTS, so the examples are just for illustration
    # and do not work.
    - id: Fedora-Rawhide-20240411.n.0
    - latest:
        release_short: Fedora
        release_version: Rawhide
        release_type: ga
        tag: nightly

packages:
  # list of rpm names to resolve
  - vim-enhanced

reinstallPackages: []
  # list of rpms already provided in the base image, but which should be
  # reinstalled

arches:
  # The list of architectures can be set in the config file. Any `--arch` option set
  # on the command line will override this list.
  - aarch64
  - x86_64

context:
    # Alternative to setting command line options. Usually you will only want
    # to include one of these options, with the exception of `flatpak` that
    # can be combined with `image` and `containerfile`
    image: registry.fedoraproject.org/fedora:latest
    containerfile: Containerfile.fedora
    flatpak: true
    bare: true
    localSystem: true
    rpmOstreeTreefile: centos-bootc/centos-bootc.yaml

High-level overview: given a list of packages, repo urls and installed packages, resolve all dependencies for the packages.

There are three options for how the installed packages can be handled.

1. Resolve in the current system (--local-system). This is probably not useful for anything.

2. Resolve in empty root (--bare, --rpm-ostree-treefile). This is useful when the final image is starting from scratch, like a base image or ostree.

   When using rpm-ostree treefile, the list of packages in input file is not needed. The tool will try to get list of required packages from the treefile. The support for some stanzas in the treefile is currently missing, so some packages may not be discovered.

3. Extract installed packages from a container image. This would be used for layered images. The base image can be explicitly provided, or discovered from Containerfile.

Getting package information from the container is tricky, and went through a few iterations:

Let’s run the solver directly in the base image. This has a few cons though:

• Solving for non-native architectures requires emulation.
• It only works if the solver is using DNF 4 and the container provides dnf. Once the solver uses DNF 5, or for any base image using microdnf (or yum, or zypper…), it doesn’t work.
  • That can not be solved by installing the solver library into the image, as it would affect the results and where would it be installed from anyway? Using a statically linked depsolver would avoid installing dependencies, but then you need a different binary for each architecture.

Let’s run the solver on the host system, but filter out base image packages after solving. Listing installed packages in the container is fairly easy, and we can rely on rpm executable being present if the user wants to install packages.

This approach doesn’t really work though.

• If the configured repos do not contain the full set of transitive dependencies, the solver will fail (or at least not see the full list of transitive dependencies, which can hide issues).
• If the configured repos contain newer versions of the packages already present on the image, the solver will include them in the result, and it becomes impossible to tell if the older version is sufficient or not.
  • If we keep the updated version in the result but the older version is fine, then we are prefetching something that will not be used.
  • If we remove the package but it is actually needed, the build process will fail.

We need to have information about the base image contents at the time the transaction is being resolved. Let’s not even consider listing package details with some incantation of rpm -qa --queryformat.

We can copy the rpmdb from the base image into some temporary directory and use that as installroot during solving. A cleaner way might be to do rpmdb --exportdb from the container and rpmdb --importdb –root into the temporary location.

So the last thing is what the tool actually implements. It will pull the image, run it and copy the rpmdb out to a temporary location on the host system. This location is used as installroot for calling DNF.

It seems to work with any architecture, though it can result in pulling quite a few images locally.

The main issue with this approach is that it's complicated to execute in containers. To be able to run podman run inside a container, the parent container has to be running with --privileged option. This may be a problem for running in CI.

An alternative would be to pull the image, mount it, and copy the rpmdb out using host tools. This way there could even be a cache for different images, avoiding some pulls. I didn't experiment with this too much. The main hurdle was turning the base image specification from Containerfile into an image id to pull. I got stuck on resolving short names, but maybe it's not necessary?

This is a minor improvement over iteration 3. The podman usage can be replaced by using skopeo, which can obtain the data witout requiring any additional permissions when running inside containers.