cluster-autoscaler-with-capi/docs/plans/2026-09-29-2315-pv1-2-workers.md

# Plan: lab1 workers on both PVE nodes (template copy per node)

**Created:** 2026-09-29 23:15 CEST

## Context

`pv1-1` is now the PVE cluster `homelab`, and `pv1-2` (192.168.0.10, 12 cores, 40.9 GiB) has joined
it. There is no shared storage, and the owner doesn't want any for now. CAPMOX v0.9.1 can't clone
across nodes from `local-lvm`, and `templateSelector` needs exactly one match in the whole cluster
([session note](../sessions/2026-09-29-2311-pve-two-node-cluster.md)). So each node needs its own
template VMID, and workers get one MachineDeployment per PVE node.

Two blockers found through the PVE API:
- `vmbr0` on `pv1-2` is not VLAN-aware.
- Template 101 exists only on `pv1-1`. Its config is two `local-lvm` disks (`scsi0`, `efidisk0`),
  no ISO or `local` references, so an offline migration is possible.

Owner's choices:
- Copy template 101 rather than rebuilding it.
- Keep `lab1-workers` as it is and add `lab1-workers-pv1-2`.

## Steps

### Step 0 — Record the plan
Copy this plan to `docs/plans/<date +%Y-%m-%d-%H%M>-pv1-2-workers.md`, create the execution log
`docs/plans-executions/<same name>` with the checklist, and commit both on their own to `main`.

### Step 1 — `vmbr0` on `pv1-2` VLAN-aware (owner runs)
In `/etc/network/interfaces` on `pv1-2`, add to the `iface vmbr0` stanza:
```
bridge-vlan-aware yes
bridge-vids 2-4094
```
Then run `ifreload -a`. The switch port for `pv1-2` must carry VLAN 4 tagged, the same as `pv1-1`.

Verify (read-only, by me): `GET /nodes/pv1-2/network/vmbr0` → `bridge_vlan_aware: 1`.

### Step 2 — Template 9101 on `pv1-2` (owner runs on `pv1-1`)
```bash
qm clone 101 9101 --full 1 --name ubuntu-2404-kube-v1.35.9 --storage local-lvm  # the clone stays on pv1-1 (local storage)
qm migrate 9101 pv1-2 --with-local-disks                                        # offline: copies both LVs to pv1-2's local-lvm
ssh root@192.168.0.10 qm template 9101                                          # on pv1-2
```
- VMID 9101 is outside CAPMOX's `nextid` range (103+), next to the existing 9000/9001 templates.
- The name stays the same on purpose. CAPMOX addresses the template by `templateID`, not by name.

Verify:
- `/cluster/resources`: 9101 is on `pv1-2` with `template: 1`.
- Smoke test from [image-builder-proxmox.md](../image-builder-proxmox.md#smoke-test-of-the-template),
  run on `pv1-2` with `TPL=9101`. It also proves that VLAN 4 works on `pv1-2` (DHCP lease, guest
  agent answers).

### Step 3 — Second worker MachineDeployment (`configs/capi/lab1.yaml`)
- New `ProxmoxMachineTemplate/lab1-worker-2g-pv1-2`: a copy of `lab1-worker-2g` with:
  - `sourceNode: pv1-2`
  - `templateID: 9101`
  - `allowedNodes: [pv1-2]` (the machine-level list overrides the cluster's,
    `vmscheduler.go:52-62`)
- New `MachineDeployment/lab1-workers-pv1-2`: a copy of `lab1-workers` with:
  - the same KubeadmConfigTemplate `lab1-worker` and the same capacity annotations
  - the same rollout `maxSurge 0 / maxUnavailable 1`
  - min 0 / max **4**
- `lab1-workers`: max-size annotation 5 → **3**. Only an annotation changes, so no rollout.
  `lab1-worker-2g` is not edited; templates are never edited in place.
- `ProxmoxCluster.allowedNodes` stays `[pv1-1]`. The control plane (template 101) must not be
  scheduled onto `pv1-2`.
- IP pool check: .31–.39 has 9 addresses; 1 CP + 3 + 4 workers = 8, so one address is spare.
- Update the header comment: generate command and hand edit #10 (second MD, per-node template),
  plus #7 for the new max-size.

Owner runs:
```bash
kubectl diff -f configs/capi/lab1.yaml   # expect: new PMT + MD, max-size 5→3 on lab1-workers
kubectl apply -f configs/capi/lab1.yaml
```
- No change to cluster-autoscaler values: `autoDiscovery` picks up the new MD, and
  `balance-similar-node-groups: true` plus the `least-waste` expander spread scale-ups across the
  two groups.
- The RBAC rule for `proxmoxmachinetemplates` already covers the new template.

### Step 4 — Scale test
1. Scale `inflate` so it needs about 4 workers (`configs/test/inflate.yaml`, 2 pods per worker).
2. Expect:
   - the CA log shows `discovered node group: MachineDeployment/default/lab1-workers-pv1-2 (min: 0, max: 4…)`
   - the scale-up is balanced across both MDs
   - the ProxmoxMachines land on `pv1-1` / `pv1-2` according to `status.proxmoxNode`
   - the nodes go Ready (Cilium)
3. Then scale down to 0 and confirm both groups shrink.
Diagnosis goes through the `cluster-inspector` agent.

### Step 5 — Docs and history
- `docs/decisions.md`:
  - new row #18: "Template copy per PVE node (101 on `pv1-1`, 9101 on `pv1-2`); one worker MD per
    node". Alternatives: shared storage, a rebuild per node, `templateSelector`. Status
    `proposed` until Step 4 passes, then `active`.
  - #7 → `superseded by #18`
- `docs/phase-1-proxmox-capi.md`:
  - topology table (two MDs, nodes)
  - Step 0: the storage/network paragraph (per-node template, VLAN-aware bridge on every node)
  - Step 5 annotations
  - troubleshooting rows:
    - clone to another node → "VM uses local storage"
    - VLAN not reachable on the new node
- `docs/image-builder-proxmox.md`: a new section "Template on other PVE nodes" with the Step 2
  commands. The result line then names both VMIDs.
- `configs/capi/clusterctl.yaml.example`: fix the comments on `ALLOWED_NODES` / `SOURCENODE`.
- CLAUDE.md, AGENTS.md "Lab facts" and README line 29: PVE cluster `homelab`, `pv1-2`
  192.168.0.10, templates 101/9101. These files have uncommitted owner edits, so I stage only my
  hunks (`git add -p`) or ask first.
- Session note: add the outcome and a link to the plan.
- `CHANGELOG.md` entry, only after the owner confirms Step 4 works.

## Verification
- PVE API (read-only): `vmbr0` VLAN-aware on both nodes; template 9101 on `pv1-2`.
- Smoke-test VM on `pv1-2` gets a VLAN 4 address.
- `clusterctl describe cluster lab1`: both MDs are healthy.
- The Step 4 scale test puts workers on both PVE nodes, and they join and become Ready.
- Rollback: the owner deletes `MachineDeployment/lab1-workers-pv1-2` and restores max 5 on
  `lab1-workers`. Template 9101 can stay.