Raspberry Pi k3s fleet — central and HA storage options¶
Parent runbook: How to provision k3s, Longhorn, and Rancher on a Raspberry Pi fleet. Conceptual compare: Longhorn vs central storage architecture.
Decision outline¶
mermaid
flowchart TD
Start[Need durable K8s volumes?] -->|Yes| LH[Longhorn on Pi fleet]
Start -->|No / edge spike| LP[local-path or hostPath P0 only]
LH --> Mix{Also large files / archive?}
Mix -->|Yes| Hybrid[Longhorn PVCs + NAS or object]
Mix -->|No| LHonly[Longhorn only]
Option A — Longhorn distributed across nodes (common)¶
Mandatory for this path (see Longhorn storage configuration sequence):
- Dedicated disk per replica node you rely on.
- open-iscsi on each schedulable Longhorn node.
- Replica count matched to reality: three replicas need three healthy disks on three nodes—not “free” on Pis.
P0/P1: Two or even one replica for non-critical PVCs may be acceptable if backups exist—document RPO honestly.
Optional (HA / scale):
- Longhorn backup target to S3-compatible storage or NFS on a central server.
- Node or disk tags to keep replicas off the etcd Pi if you want control-plane separation (validate against current Longhorn docs).
Option B — Central NAS / NFS (or SMB) for specific workloads¶
Use for: large media, read-mostly archives, backup landing zones.
Mandatory:
- Network path stable enough for your app’s latency and loss budget.
- Avoid starving the etcd disk with synchronous NFS-heavy traffic on the same Pi without measurement.
Optional (HA / scale): dual NAS heads or off-site rsync—later.
Option C — Multiple storage servers (hybrid)¶
Pattern: Two or more Pis (or small x86 hosts) with good disks as Longhorn nodes, plus a NAS used only as a backup target.
P1-friendly: Keeps Kubernetes PVCs in Longhorn while offloading bulk files to central storage.
k3s control-plane HA (storage-adjacent)¶
| Pattern | Phase | Notes |
|---|---|---|
| Single embedded etcd server | P0/P1 | Simplest; backup cluster state per k3s docs and your install method. |
| Three or more HA servers | Later | Follow k3s HA embedded etcd exactly—test failure in staging first. |
Rule: Storage HA (Longhorn replicas) does not replace control-plane HA or off-site backups.