Architecture¶

raya can be divided into three layers:

The platform. raya is made of several VMs, one running k3s’ control plane and as many agents as declared in the repo (see deploy/kube-system/agents.json). They are hosted by Hetzner, run a Fedora CoreOS system provisioned via Ignition, and communicate via a private network. A lot of efforts have been dedicated to ensure the VMs themselves can be thrown away.

The cluster services. I have tried to build a convenient production environment for my experiments. I kept k3s’ choice for the Ingress controller (Traefik), set up external-dns and cert-manager, provided the necessary credentials to Hetzner CSI driver to dynamically allocate external disks, provisioned a shared Postgres instance with WAL archiving and nightly backups to a Cloudflare bucket using CloudNativePG, and automated my deployments from Git repositories using Flux.

The workloads. The actual services I have deployed on top of raya are defined in a dedicated repository. They are not documented here.

For each layer, the same guiding principle applies: manual actions should be limited to the strict minimum. Deployment happens automatically after a change reaches the main branch of the appropriate repositories, either because the CI runs the deployment or because the cluster itself applies the changes.