Skip to main content

An experimental harness tool based on systemd-nspawn containers

Project description

Operating system containers for humans and machines.


systemd System and Service Manager Python version Version Maturity status Test suite status Test suite code coverage License PyPI downloads

About

An experimental harness tool based on systemd and systemd-nspawn, to run operating system containers, in the spirit of addressing some details of Docker Considered Harmful and Systemd vs. Docker.

At the same time, it is a tribute to the authors and contributors of GNU, Linux, systemd, Python, VirtualBox, Vagrant, Docker, and more.

Most people running Linux probably want to use Podman these days. For more background, enjoy reading Container wars and Container Tools Guide.

Racker is …

  • A runtime harness for testing software packages and similar purposes, in different environments, mostly run headless and non-interactively.

  • A lightweight wrapper around systemd-nspawn to provide and launch container environments for/with systemd.

  • A lightweight wrapper around vagrant to provide convenient access to all things needing a full VM, like running Windows on Linux or macOS.

Comparison with similar tools

The aims of Racker are very similar to Distrobox and Toolbox. However, there are also some differences.

  • Racker is currently based on systemd-nspawn instead of Docker or Podman.

  • Racker can invoke any kind of container payload, but strongly focuses on running OS containers aka. OS-level virtualization, using systemd as init process.

  • Racker aims to provide concise usability by folding its usage into a single command.

  • The acquisition and provisioning of operating system images does not need any special preparation steps, those are handled by Racker on the fly.

See also Comparison with similar tools - more details.

About systemd-nspawn

As a general introduction, we recommend to read an introductory article on LWN as well as two installments of the systemd for administrators blog series:

  • Creating containers with systemd-nspawn

    • Lennart Poettering spoke about a mostly unknown utility that ships with it: systemd-nspawn. The tool started as a debugging aid for systemd development, but has many more uses than just that, he said. […]

    • The idea was to write a tool that does much of what LXC and libvirt LXC do, but is easier to use. It is targeted at “building, testing, debugging, and profiling”, not at deployment. systemd-nspawn uses the same kernel APIs that the other two tools use, but is not a competitor to them because it is not targeted at running in a production environment. […]

  • Changing Roots

    • As administrator or developer sooner or later you’ll encounter chroot() environments. […]

    • File system namespaces are in fact a better replacement for chroot() in many many ways. […]

    • More importantly however systemd comes out-of-the-box with the systemd-nspawn tool which acts as chroot on steroids: it makes use of file system and PID namespaces to boot a simple lightweight container on a file system tree. […]

  • OS containers

    • We’ll focus on OS containers here, i.e. the case where an init system runs inside the container, and the container hence in most ways appears like an independent system of its own. […]

    • We use systemd-nspawn extensively when developing systemd. […]

Lennart Poettering, the author of systemd, identifies three main pillars of containers [1]:

  • Resource bundling

  • Sandboxing

  • Delivery

At [2] Lennart Poettering and Kai Sievers outline their vision of systemd as a platform for running systems and their focus on containers in 2014. Fast forward to 2022, and everything is pretty much there. systemd now provides a plethora of features for containerization, specifically for resource bundling and sandboxing [1].

[3] outlines how systemd-nspawn was originally conceived to aid in testing and debugging systemd, [4] is the latest overview of systemd in 2018. For approaching systemd-nspawn from a user’s perspective, a concise introductory walkthrough can be found at [5].

The most important bits being covered by the systemd software family already, Racker tries to fill some gaps on the delivery aspects.

Setup

Install prerequisites:

apt-get update
apt-get install --yes systemd-container skopeo umoci python3-pip python3-venv

Install Racker:

python3 -m venv .venv
source .venv/bin/activate
pip install racker --upgrade

To install the latest development version, use this command instead:

pip install git+https://github.com/cicerops/racker --upgrade

Usage

Racker

The racker program aims to resemble the semantics of Docker by providing a command line interface compatible with the docker command.

# Invoke the vanilla Docker `hello-world` image.
# FIXME: Does not work yet.
# racker run -it --rm hello-world /hello
# racker run -it --rm quay.io/podman/hello

# Acquire rootfs images.
racker pull debian:bullseye-slim
racker pull fedora:37

# Launch an interactive shell.
racker run -it --rm debian:bullseye-slim bash
racker run -it --rm fedora:37 bash
racker run -it --rm docker://registry.access.redhat.com/ubi8/ubi-minimal /bin/bash
racker run -it --rm docker://quay.io/centos/centos:stream9 bash

# Launch a single command.
racker run -it --rm debian:11-slim hostnamectl
racker run -it --rm opensuse/tumbleweed hostnamectl
racker run -it --rm ubuntu:jammy /bin/cat /etc/os-release
racker run -it --rm registry.suse.com/suse/sle15 /bin/cat /etc/os-release
racker run -it --rm registry.suse.com/bci/bci-base:15.4 /bin/cat /etc/os-release
racker run -it --rm docker://ghcr.io/jpmens/mqttwarn-standard /usr/bin/hostnamectl

# Verbose mode.
racker --verbose run -it --rm fedora:37 hostnamectl

# Use stdin and stdout, with timing.
time echo "hello world" | racker run -it --rm fedora:37 cat /dev/stdin > hello
cat hello

Postroj

The idea behind postroj is to provide an entrypoint to a command line interface implementing actions that don’t fit into racker, mostly having a more high-level character.

Currently, postroj pkgprobe implements a flavor of full system integration/acceptance testing in order to test the soundness of actual installed binary distribution packages, in the spirit of autopkgtest.

To do so, it implements the concept of curated operating system images, whose labels have a different layout than labels of Docker filesystem images.

Getting started:

# List available images.
postroj list-images

# Acquire images for curated operating systems.
postroj pull debian-bullseye
postroj pull fedora-37

# Acquire rootfs images for all available distributions.
postroj pull --all

# Run a self test procedure, invoking `hostnamectl` on all containers.
postroj selftest hostnamectl

Package testing:

# Run a self test procedure, invoking example probes on all containers.
postroj selftest pkgprobe

# Run two basic probes on different operating systems.
postroj pkgprobe --image=debian-bullseye --check-unit=systemd-journald
postroj pkgprobe --image=fedora-37 --check-unit=systemd-journald
postroj pkgprobe --image=archlinux-20220501 --check-unit=systemd-journald

# Run probes that need to install a 3rd party package beforehand.

postroj pkgprobe \
    --image=debian-stretch \
    --package=http://ftp.debian.org/debian/pool/main/w/webfs/webfs_1.21+ds1-12_amd64.deb \
    --check-unit=webfs \
    --check-network=http://localhost:8000

postroj pkgprobe \
    --image=debian-bullseye \
    --package=https://dl.grafana.com/oss/release/grafana_8.5.1_amd64.deb \
    --check-unit=grafana-server \
    --check-network=http://localhost:3000

postroj pkgprobe \
    --image=centos-8 \
    --package=https://dl.grafana.com/oss/release/grafana-8.5.1-1.x86_64.rpm \
    --check-unit=grafana-server \
    --check-network=http://localhost:3000

Performance

A SuT which just uses a dummy probe /bin/systemctl is-active systemd-journald on Debian 10 “buster” cycles quite fast, essentially demonstrating that the overhead of environment setup/teardown is insignificant.

time postroj pkgprobe --image=debian-buster --check-unit=systemd-journald

real    0m0.589s
user    0m0.161s
sys     0m0.065s

On a cold system, where the filesystem image would need to be acquired before spawning the container, it’s still fast enough:

time postroj pkgprobe --image=debian-bookworm --check-unit=systemd-journald

real    0m22.582s
user    0m8.572s
sys     0m3.136s

Questions and answers

  • Q: How does it work?
    A: Directly quoting the machinectl documentation here:

    Note that systemd-run with its --machine= switch may be used in place of the machinectl shell command, and allows non-interactive operation, more detailed and low-level configuration of the invoked unit, as well as access to runtime and exit code/status information of the invoked shell process.

    In particular, use systemd-run’s --wait switch to propagate exit status information of the invoked process. Use systemd-run’s --pty switch for acquiring an interactive shell, similar to machinectl shell. In general, systemd-run is preferable for scripting purposes.

  • Q: How does it work, really?
    A: Roughly speaking…
    • skopeo and umoci are used to acquire root filesystem images from Docker image registries.

    • systemd-nspawn is used to run commands on root filesystems for provisioning them.

    • Containers are started with systemd-nspawn --boot.

    • systemd-run is used to interact with running containers.

    • machinectl is used to terminate containers.

  • Q: How is this project related with Docker?
    A: The runtime is completely independent of Docker, it is solely based on systemd-nspawn containers instead. However, root filesystem images can be pulled from Docker image registries in the spirit of machinectl pull-dkr. Other than this, the racker command aims to be a drop-in replacement for its corresponding docker counterpart.
  • Q: Do I need to have Docker installed on my machine?
    A: No, Racker works without Docker.
  • Q: How are machine names assigned?
    A: Machine names for spawned containers are automatically assigned. The name will be assembled from the distribution’s fullname attribute, prefixed with postroj-. Examples: postroj-debian-buster, postroj-centos-8.
  • Q: Does the program need root privileges?
    A: Yes, the program currently must be invoked with root or corresponding sudo privileges. However, it would be sweet to enable unprivileged operations soon. systemd-nspawn should be able to do it, using --private-users or --user?
  • Q: Where does the program store its data?
    A: Data is stored at /var/lib/postroj. In this manner, it completely gets out of the way of any other images, for example located at /var/lib/machines. Thus, any images created or managed by Racker will not be listed by machinectl list-images.
    A: The download cache is located at /var/cache/postroj/downloads.
  • Q: Where are the filesystem images stored?
    A: Activated filesystem images are located at /var/lib/postroj/images.
  • Q: How large are curated filesystem images?
    A: The preference for curated filesystem images is to use their corresponding “slim” variants where possible, aiming to only use artefacts with download sizes < 100 MB.
  • Q: Are container disks ephemeral?
    A: Yes, by default, all container images will be ephemeral, i.e. all changes to them are volatile.

Troubleshooting

It’s always the cable. ;]

  1. If you see that your container might not have network access, make sure to provide a valid DNS configuration in your host’s /etc/resolv.conf. When in doubt, please add nameserver 9.9.9.9 as the first entry.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

racker-0.2.0.tar.gz (74.7 kB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page