Harvester hardware planning

11.1 Minimums vs. reality¶

	Testing minimum	Production recommendation
CPU	x86_64 with VT-x/AMD-V, 8 cores	16+ cores per node (VMs need room)
RAM	32 GB	64 GB+ (128–512 GB typical)
Boot/OS disk	250 GB SSD	250–500 GB SSD/NVMe, separate from data
Data disks	Can share OS disk in tests	Dedicated NVMe/SSD, multiple per node
Network	1 GbE	2× 10/25 GbE bonded; separate mgmt/VM/storage if possible
Nodes	1 (no HA)	3+ (or 2 + witness)

ARM64 support has arrived in recent releases as a technical preview — check the release notes before planning ARM hardware.

Disk performance: HDDs are explicitly not recommended. Longhorn + etcd on spinning rust leads to etcd latency warnings and VM I/O stalls.
etcd on the OS disk: the OS disk hosts etcd for control-plane nodes; NVMe here matters more than people expect.
RAM headroom: Harvester's own system pods (Rancher, Longhorn, KubeVirt, monitoring) consume a meaningful amount of memory per node — budget ~16–20 GB of overhead on each management node before VMs.
Uniformity: mixed CPU generations work, but live migration needs compatible CPU models — use a common CPU model in VM specs or node labelling (host-model ties VMs to similar hosts).
NIC naming: install with the final NIC layout; renaming/re-cabling later means editing network configs.
BIOS/UEFI: enable virtualization extensions, IOMMU (VT-d/AMD-Vi) for passthrough, and SR-IOV if needed. Disable Secure Boot only if your version doesn't support it.

Management (control-plane) nodes: the first three nodes become control plane + etcd automatically. Additional nodes join as workers, and Harvester promotes workers if a management node is lost (configurable).
Witness node (since 1.3): runs only etcd to give a 2-node cluster quorum. Cheap hardware is fine, but it must be reliable and low-latency to the others.
Worker-only nodes: can be explicitly installed as worker role.
Single node: supported for edge/test; no HA, Longhorn replica count must be 1.