Skip to content

0005. One node, and what that gives up

Status: Accepted, 2026-08-15

Context

Version 1 ran on a Dell PowerEdge. It was retired partly because a full rack is loud, hot and expensive to run. Version 2 went to a single Optiplex explicitly because UK electricity hit 68.3p/kWh.

That constraint has largely gone. The Ofgem cap for July to September 2026 is 26.11p/kWh, and an Optiplex idling around 12W costs roughly £25 to £38 a year. So the reason for one node is no longer cost.

Decision

Stay on one node anyway, and be explicit in the documentation about what that costs rather than pretending the cluster is more resilient than it is.

Consequences

The things that genuinely do not work, and are documented wherever they matter:

  • Every upgrade is an outage. Cordon and drain have nowhere to send anything. The API server goes down with the node it runs on.
  • PodDisruptionBudgets protect nothing. Talos ignores eviction failures during its own drains (an open upstream issue), and there would be nowhere to evict to regardless. A minAvailable: 1 PDB here is a false sense of safety.
  • A naive upgrade can strand the node, because the pod running the upgrade gets evicted before the reboot finishes. tuppr handles this by skipping drain entirely on single-node clusters, which is why it is the planned upgrade mechanism.
  • etcd has one member. There is no quorum to lose, but there is also no peer to recover from. talosctl bootstrap --recover-from assumes a surviving node, which does not exist here. An etcd snapshot covers corruption on a live host, not a dead disk.
  • Charts assume more nodes. Default affinity blocks and topology spread constraints leave replicas Pending forever. Every chart's defaults get checked.

What is deliberately not attempted: distributed storage, HA control plane, multi-replica anything, Alertmanager clustering.

Because the cluster's own monitoring dies with the node, the deadman switch has to live outside it. The Watchdog alert is routed to healthchecks.io from a Pi Zero W.

Revisiting

A Minisforum MS-03 class machine is expected within months. That is a rebuild with a fresh etcd rather than a move, so every persistent volume has to return from a backup. This is why backups are a prerequisite for the hardware upgrade and not an optional extra.