01 · Overview & management plane
Production Architecture · v1.0 — Harvester HCI: VM, Longhorn & External Rook/Ceph Storage
18-node Harvester (RKE2) · application RKE2 cluster · dedicated storage RKE2 cluster running Rook/Ceph · isolated pod & storage networks · every VM on its own host · MetalLB + Gateway API ingress. · interactive original ↗
The seven networks above are the colour key used by every diagram in this section; 03 · Network fabric describes each one.
Management plane
Rancher MCM
HA Rancher (3 replicas) on the application RKE2 cluster or a dedicated mgmt cluster. Manages Harvester + both Level-1 clusters via the Harvester node driver.
Cluster access
Harvester VIP · 10.10.100.10 — kube-vip on the mgmt network fronts the Harvester API/UI. App-cluster APIs get their own VIPs on guest VLANs.
Out-of-band
BMC / IPMI · VLAN 90 — physically isolated OOB network for lights-out node management, firmware, console. Never routed to workload networks.
Production ground rules
- Separation. Pod overlay on bond0; Longhorn + Ceph on bond1. Storage rebuild storms cannot starve workloads.
- Quorum. Every cluster: 3 masters + ≥3 workers. Harvester, application RKE2, storage RKE2, Ceph MONs — masters spread across all 3 zones.
- Disk ownership. Longhorn owns SSDs; Ceph owns passthrough NVMe. Never layer Ceph OSDs on Longhorn volumes.
- Blast radius. Storage RKE2 is a separate cluster: Ceph upgrades/failures don't touch the workload cluster, and vice versa.
- MTU. Jumbo frames (9000) end-to-end on VLANs 200/210/211; validate with
ping -M dobefore go-live. - Capacity. N-1 host failure headroom for VM restart; Ceph sized for host-out recovery without hitting full ratios.
- Backups. Longhorn backup target + Ceph RGW/off-site — snapshots are not backups.
- OOB isolation. BMC VLAN 90 reachable only from a hardened jump host; no route to workload networks.