Skip to content

Platform decision memo — phase, HA scope, deferrals (Pi / k3s / Longhorn / Rancher)

Purpose

Decision memo for the smart-farm-wiki homelab/edge Kubernetes platform: what belongs in Phase 0/1, what is deferred, what “HA” means in this repo, and what must not be overbuilt before workload truth (farmOS, telemetry) exists.

Doctrine package (first-class hub): Platform doctrine package — homelab / farm edge.

Canonical stack doc: Platform strategy for farm and homestead services.

HA semantics (expanded): HA meaning and constraints — homelab / farm platform. Component roles: k3s role, Longhorn role.

Phase 0/1 (pilot-ready)

In scope

  • 1–3 Pi-class nodes: one k3s server + agents per k3s quick-start—embedded datastore acceptable for non-life-safety lab use.
  • Longhorn trial only if iSCSI + disk plan are ready (Longhorn CSI on K3s); otherwise local-path or hostPath spikes (throwaway).
  • One stateful pilot: farmOS or broker + DB—with logical backup proven (Backup strategy comparison).
  • Documentation: tokens, kubeconfig hygiene, hostname uniqueness (Pi fleet provisioning standard).

Out of scope for 0/1

  • Rancher by default (Rancher role and timing).
  • Multi-server etcd HA without a written restore drill.
  • GitOps multi-cluster unless you already have more than one cluster to operate.

Deferred (until justified)

Item Defer until
k3s HA control plane (3+ servers) Uptime need exceeds maintenance windows you can tolerate; read HA embedded etcd and test failover
Rancher More than one cluster to operate, or explicit need for catalog / CD (Rancher role and timing)
Longhorn as default for every workload I/O and CPU cost understood on Pi; central NAS may win for bulk (Longhorn vs central storage)
Service mesh, complex ingress tiers Ten microservices you do not have yet

What “HA” means in this repository

Term Meaning Not implied
k3s HA Multiple server nodes + quorum datastore per k3s docs (architecture) Multi-region; zero RPO for farm records without app-level backup
Longhorn replication Survive node/disk loss on the same farm LAN Geographic DR; ransomware immutability (needs object-lock / air-gap policy)
Central NAS “HA” Vendor appliance feature set Kubernetes SLO unless integrated as PV backend

Full doctrine: HA meaning and constraints — homelab / farm platform.

Anti-patterns (do not overbuild)

  1. Kubernetes for a single static container on one Pi—use Docker Compose first when appropriate (farmOS Docker capture is a lighter-weight pattern for dev).
  2. Longhorn before stable k3s + disk + iSCSI—debug the base cluster first (CSI on K3s).
  3. Rancher before you can restore etcd / PVC—the dashboard is not a backup.