Design: a native platform: proxmox¶
Shiftmox gets OpenShift onto Proxmox with what the installer already has:
baremetal (agent or IPI) plus a Redfish BMC, or
external plus the Proxmox CCM. A real
platform: proxmox — like vsphere or nutanix — doesn't exist. This page
describes what it would take, so the work can be planned as a project of its
own.
What a platform is in OpenShift¶
| Piece | Repository | For Proxmox |
|---|---|---|
| install-config types + validation | openshift/installer pkg/types/proxmox |
API URL, token, node(s), storage, bridge, template |
| infrastructure provisioning | installer, Cluster API provider (since 4.16 the installer runs CAPI in-process) | CAPMOX creates the bootstrap + master VMs |
Infrastructure CR |
openshift/api (PlatformType, ProxmoxPlatformSpec/Status) |
new enum value + spec — an API change in the release payload |
| Machine API provider | machine-api-provider-proxmox (new) |
MachineSets → VMs; could wrap CAPMOX or the Proxmox API directly |
| cloud-controller-manager | cluster-cloud-controller-manager-operator |
ship the Proxmox CCM as an operand |
| storage | cluster-storage-operator |
ship the Proxmox CSI driver + operator |
| on-prem networking | machine-config-operator templates |
keepalived/haproxy/coredns as on vSphere/Nutanix |
| boot image | RHCOS/SCOS qcow2 imported as a Proxmox template | per-release image stream metadata |
| credentials | cloud-credential-operator |
Proxmox token in kube-system |
Effort, roughly¶
- Fork installer + API (types, validation, assets, CAPMOX in-process,
bootstrap ignition via cloud-init/
fw_cfg): the bulk of the work. - Machine API provider: a controller turning
Machineobjects into VMs (clone template, set cloud-init user-data = worker ignition, start). - Operators: wire CCM + CSI into the existing operators, or install them
as extra manifests (what
externalalready does). - Release payload: none of the above is in Red Hat's or OKD's payload.
It means building and signing a custom OKD release (
oc adm release new) and keeping it current with every z-stream. For OpenShift (ocp) this is not supportable.
Realistically: weeks of Go work for a first version, and an ongoing rebase burden. OKD would be the only target.
Why Shiftmox doesn't do this (yet)¶
The three install methods cover most of what a native platform adds:
| Native-platform feature | Shiftmox today |
|---|---|
| installer creates all VMs | ipi (installer powers + installs, Ansible only creates empty VMs) |
| MachineSets that create VMs | Metal3 + Redfish BMC on-demand slots (agent, ipi) |
| providerID, zones, node lifecycle | Proxmox CCM (external) |
| dynamic volumes on Proxmox storage | Proxmox CSI (all methods) |
The missing combination is Machine API and CCM at the same time: platform
baremetal reserves the providerID for Metal3, platform external has no
Machine API. A possible middle step, without forking the installer, is a small
Machine API provider for platform External that reuses the Redfish driver's VM
creation — a project for later.