Install methods¶
install_method in group_vars/all/cluster.yml decides how Shiftmox installs
the cluster. All three work for OpenShift and OKD (ocp_distribution).
agent (default) |
ipi |
external |
|
|---|---|---|---|
| Installer | agent-based (openshift-install agent) |
baremetal IPI (openshift-baremetal-install) |
agent-based |
| Platform | baremetal (none for SNO) |
baremetal |
external (platformName: proxmox) |
| Who creates the VMs | Ansible | Ansible, empty; Metal3 installs them | Ansible |
| How nodes boot the installer | agent ISO in the CD drive | Redfish virtual media, per node | agent ISO in the CD drive |
| Bootstrap | on the first master (rendezvous) | temporary VM on a provisioner VM | on the first master |
| API / ingress | keepalived VIPs | keepalived VIPs | HAProxy on the runner host (external_lb_address) |
| Masters and workers as BareMetalHosts | no (autoscale workers only) | yes, all of them | no |
Worker autoscaling (./ocp.sh autoscale) |
yes | yes | no Machine API |
| Proxmox cloud-controller-manager | no | no | yes (providerID, zones, node lifecycle) |
| Proxmox CSI | optional | optional | optional |
| Topologies | SNO, compact, standard | compact, standard (3 masters) | SNO, compact, standard |
| Status | tested | experimental | experimental |
Which one?¶
agent— the proven path. Fewest moving parts, no extra VMs.ipi— everything is managed by Metal3 from the first minute: the console shows every node as a bare metal host with power controls, and replacing a broken master or worker goes through the normal OpenShift machine flows. Needs a provisioner VM with nested virtualization.external— Kubernetes knows the nodes are Proxmox VMs: nodes getproviderID: proxmox://<region>/<vmid>andtopology.kubernetes.io/zone= Proxmox node, and a Node whose VM is deleted disappears from the cluster. The price is no Machine API (no MachineSets, no autoscaling) and an external load balancer.
A real platform: proxmox in the installer doesn't exist; the
design notes describe what it would take.
Switching¶
The method is fixed when the cluster is installed. To switch:
./ocp.sh destroy, change install_method, ./ocp.sh deploy.