Kubernetes on a small node

k3s

k3s is a complete Kubernetes distribution in a single binary, aimed at small machines and edge devices. One command installs a single-node cluster with Traefik (ingress), CoreDNS, a local storage provisioner and a metrics server already running:

curl -sfL https://get.k3s.io | sh -
sudo k3s kubectl get nodes
sudo /usr/local/bin/k3s-uninstall.sh   # removes it again, cleanly

What it costs

The control plane (API server, scheduler, controller manager, embedded datastore) plus the bundled system pods use roughly 600 to 800 MB before any of your own workloads run. Add-ons cost more: cert-manager alone is three more pods. The k3s docs list 2 GB as a sensible minimum for a server node; 4 GB is comfortable.

How a 1 GB node fails

Not with a clean error, but by slowing down until things time out:

• kubectl reports TLS handshake timeout or request did not complete within requested timeout.
• Installing manifests times out half-way, leaving a broken half-installed add-on.
• Finally SSH stops answering too, because the kernel is busy swapping and killing processes.

Checks while you still can get in: free -h, uptime (load average), sudo journalctl -u k3s -n 50, dmesg | grep -i -E 'oom|killed'. Recovery from outside: restart the VM from the cloud console.

Lessons for small clusters

• Do the memory arithmetic before choosing the platform.
• Waits in setup scripts need a timeout and a message; "the node exists" comes before "the node is Ready"; CRDs must exist before resources that use them (no matches for kind ... means they don't yet).
• One heavy add-on at a time, and remove what failed before retrying.

Alternatives

• Docker Compose for one machine (Docker Compose on a small VM).
• A bigger node for the cluster, for example Oracle Cloud's free Arm instance, and keep the manifests ready for that day.
• For learning only: kind or minikube on a laptop, where memory is plentiful.