28.1 Edge / branch (2 nodes + witness)¶

[node1: 16c/128G/2×NVMe]   [node2: 16c/128G/2×NVMe]   [witness: small/NUC]
         └────────── 2×10G LACP, VLANs: mgmt, vm, storage ──────────┘
Longhorn replicas: 2 · Backups → central S3 · Managed by central Rancher

28.2 Datacenter HCI (6–20 nodes)¶

• 3 management + N worker nodes, spread across racks (topology.kubernetes.io/zone labels; Longhorn replica zone anti-affinity).
• Dedicated storage network, 25G, MTU 9000.
• External S3 backup target in another failure domain.
• Rancher on its own 3-node RKE2 cluster (could itself be VMs on a different Harvester cluster — avoid circular dependency on the same one).
• Guest RKE2 clusters per environment/tenant.

28.3 Multi-site¶

• One Harvester cluster per site (don't stretch a cluster across a WAN — etcd and Longhorn replication need low latency).
• Central Rancher manages all sites; Fleet pushes images, networks, policies, guest-cluster definitions.
• Cross-site DR via shared backup targets and periodic restore testing.

28.4 Sizing worked example¶

Target: 120 VMs averaging 4 vCPU / 8 GiB / 80 GiB.

• vCPU: 480 vCPU. At 4:1 CPU overcommit → 120 physical cores worth of threads → with 32-core/64-thread nodes, ~4 nodes for CPU (plus overhead).
• RAM: 960 GiB for guests at 1:1 + ~20 GiB/node system overhead. With 384 GiB nodes → 3 nodes for RAM, N+1 → 4, with headroom for a node failure (evacuate 1/4 of load) → 5 nodes.
• Storage: 9.6 TB provisioned, assume 60% actually used thin → ~5.8 TB × 3 replicas = 17.3 TB; at 70% target utilization → ~25 TB raw → 5 nodes × 2 × 3.84 TB NVMe ≈ 38 TB raw (room to grow).
• Result: 5 nodes, 32c/384G/2× 3.84 TB NVMe + boot SSD, 2×25G data + 2×10G mgmt.

← Migrating from VMware or OpenStack · Index · Homelab path →