GitHub runner

  • Canonical IS DevOps
Channel Revision Published Runs on
latest/stable 308 11 Dec 2024
Ubuntu 22.04 Ubuntu 20.04
latest/stable 302 02 Dec 2024
Ubuntu 22.04 Ubuntu 20.04
latest/stable 1 09 Feb 2022
Ubuntu 22.04 Ubuntu 20.04
latest/beta 317 11 Dec 2024
Ubuntu 22.04
latest/beta 310 28 Nov 2024
Ubuntu 22.04
latest/edge 322 20 Dec 2024
Ubuntu 22.04 Ubuntu 20.04
latest/edge 310 28 Nov 2024
Ubuntu 22.04 Ubuntu 20.04
latest/edge 4 26 Apr 2022
Ubuntu 22.04 Ubuntu 20.04
local-lxd/stable 306 17 Dec 2024
Ubuntu 22.04
local-lxd/edge 306 27 Nov 2024
Ubuntu 22.04
1/stable 177 05 Jun 2024
Ubuntu 22.04
1/edge 177 05 Jun 2024
Ubuntu 22.04
juju deploy github-runner
Show information

Platform:

Ubuntu
22.04 20.04

Charm architecture

A Juju charm to operate a set of GitHub self-hosted runners while managing security and resource usage.

Conceptually, the charm can be divided into the following:

LXD ephemeral virtual machines

To ensure a clean and isolated environment for every runner, self-hosted runners use LXD virtual machines. The charm spawns virtual machines, setting resources based on charm configurations. The self-hosted runners start with the ephemeral option and will clean themselves up once the execution has finished, freeing the resources. This is similar to how GitHub hosts their runners due to security concerns.

As the virtual machines are single-use, the charm will replenish virtual machines on a regular schedule. This time period is determined by the reconcile-interval configuration.

On schedule or upon configuration change, the charm performs a reconcile to ensure the number of runners managed by the charm matches the virtual-machines configuration, and the resources used by the runners match the various resource configurations.

To prevent disk IO exhaustion on the Juju machine on disk-intensive GitHub workflows, the charm has two storage options provided by runner-storage configuration:

  • Use memory of the Juju machine as disk. Under this option, the vm-disk configuration can impact the memory usage of the Juju machine.
  • Use storage mount by Juju as the disk.

Virtual machine image

The virtual machine images are built on installation and on a schedule every 6 hours. These images are constructed by launching a virtual machine instance, modifying the instance with configurations and software installs, and then exporting the instance as an image. This process reduces the time needed to launch a virtual machine instance for hosting the self-hosted runner application.

The software installed in the image includes:

  • APT packages:
    • docker.io
    • npm
    • python3-pip
    • shellcheck
    • jq
    • wget
  • npm packages:
    • yarn
  • Binary downloaded:
    • yq

The configurations applied in the image include:

  • Creating a group named microk8s.
  • Adding the ubuntu user to the microk8s group. Note that the microk8s package is not installed in the image; this preconfigures the group for users who install the package.
  • Adding the ubuntu user to the docker group.
  • Adding iptables rules to accept traffic for the DOCKER-USER chain. This resolves a networking conflict with LXD.

Network configuration

The charm respects the HTTP(S) proxy configuration of the model configuration of Juju. The configuration can be set with juju model-config using the following keys: juju-http-proxy, juju-https-proxy, juju-no-proxy. The GitHub self-hosted runner applications will be configured to utilise the proxy configuration. This involves setting environment variables such as http_proxy, https_proxy, no_proxy, HTTP_PROXY, HTTPS_PROXY, and NO_PROXY in various locations within the runner environment, such as /etc/environment.

However, employing this approach with environment variables has its drawbacks. Not all applications within a workflow may adhere to these variables as they lack standardisation. This inconsistency can result in failed workflows, prompting the introduction of aproxy, as detailed in the subsection below.

aproxy

If the proxy configuration is utilised and aproxy is specified through the charm’s configuration option, all HTTP(S) requests to standard ports (80, 443) within the GitHub workflow will be automatically directed to the specified HTTP(s) proxy. Network traffic destined for ports 80 and 443 is redirected to aproxy using iptables. aproxy then forwards received packets to the designated HTTP(S) proxy. Beyond that, the environment variables (http_proxy, https_proxy, no_proxy, HTTP_PROXY, HTTPS_PROXY, NO_PROXY) will no longer be defined in the runner environment. It’s worth noting that this setup deviates from the behaviour when not using aproxy, where these variables are set in the runner environment. In that scenario, traffic to non-standard ports would also be directed to the HTTP(s) proxy, unlike when using aproxy.

denylist

The nftables on the Juju machine are configured to deny traffic from the runner virtual machine to IPs on the denylist configuration. The runner will always have access to essential services such as DHCP and DNS, regardless of the denylist configuration.

GitHub API usage

The charm requires a GitHub personal access token for the token configuration. This token is used for:

  • Requesting self-hosted runner registration tokens
  • Requesting self-hosted runner removal tokens
  • Requesting a list of runner applications
  • Requesting a list of self-hosted runners configured in an organization or repository
  • Deleting self-hosted runners

The token is also passed to repo-policy-compliance to access GitHub API for the service.

Note that the GitHub API uses a rate-limiting mechanism. When this is reached, the charm may not be able to perform the necessary operations and may go into BlockedStatus. The charm will automatically recover from this state once the rate limit is reset, but using a different token with a higher rate limit may be a better solution depending on your deployment requirements.

GitHub repository setting check

The repo-policy-compliance is a Flask application hosted on Gunicorn that provides a RESTful HTTP API to check the settings of GitHub repositories. This ensures the GitHub repository settings do not allow the execution of code not reviewed by maintainers on the self-hosted runners.

Using the pre-job script, the self-hosted runners call the Python web service to check if the GitHub repository settings for the job are compliant. If not compliant, it will output an error message and force stop the runner to prevent code from being executed.

Dependencies management

Upon installing or upgrading the charm, the kernel will be upgraded, and the Juju machine will be restarted if needed.

The charm installs the following dependencies:

  • For running repo-policy-compliance
    • gunicorn
  • For firewall to prevent runners from accessing web service on the denylist
    • nftables
  • For virtualization and virtual machine management
    • lxd
    • cpu-checker
    • libvirt-clients
    • libvirt-daemon-driver-qemu
    • apparmor-utils

These dependencies can be regularly updated using the landscape-client charm.

The charm installs the following dependencies and regularly updates them:

  • repo-policy-compliance
  • GitHub self-hosted runner application

The charm checks if the installed versions are the latest and performs upgrades if needed before creating new virtual machines for runners.

COS Integration

Upon integration through the cos-agent, the charm initiates the logging of specific metric events into the file /var/log/github-runner-metrics.log. For comprehensive details, please refer to the pertinent specification.

Subsequently, the grafana-agent transmits this log file to Loki, facilitating access for Grafana’s visualisation capabilities. Notably, most events are transmitted during reconciliation. This approach prioritises long-term monitoring over real-time updates, aligning with the intended monitoring objectives.


Help improve this document in the forum (guidelines). Last updated 2 months ago.