A pragmatic learning progression:
- Single-node Harvester on a spare box (32 GB+ RAM, NVMe). Learn images, VMs, VLANs (a managed switch with 802.1Q helps), cloud-init.
- Nested lab: three Harvester VMs on a KVM host with nested virtualization (
kvm_intel nested=1), 16–24 GB RAM each — slow, but exercises HA, live migration, maintenance mode and upgrades without three physical machines. - Rancher on a small K3s cluster; import Harvester; build an RKE2 guest cluster with the Harvester CCM/CSI.
- Terraform everything (images, networks, VMs) and keep it in Git; then try Fleet for images and networks.
- Backups to MinIO or an S3-compatible bucket; practice a full restore into a second (single-node) Harvester.
- Upgrade across a patch release and watch the A/B flow.
- Break things: pull a power cable mid-write, fill a disk, misconfigure a VLAN — and recover.
Hardware notes for labs: Harvester won't be happy below 32 GB RAM per node; consumer NVMe works but watch endurance under Longhorn replication; 2.5G NICs work for learning but 10G makes a visible difference.
← Reference architectures · Index · Gotchas and hard-won lessons →