Placement and disks

affinity_host steers each VM to a physical host, but the real anchor is the host-local SR its disks sit on: a VM cannot boot where its storage is not.

Node	Host	SR
talos-01	gewisvhost1	vhost1-ssd2
talos-02	gewisvhost3	vhost3-ssd
talos-03	gewisvhost4	Local storage

One node per host, so losing a host costs one etcd member and one Longhorn replica. The pool is vhost1/vhost3/vhost4 — there is no vhost2. On vhost4, vhost4-ssd is nearly full; Local storage is the SR with room.

Each node gets 4 vCPU and static memory set per node in the nodes map: 30 GiB, except 16 GiB for talos-01 because vhost1 has no more room until the old cluster's swarm201 is gone. Raising it is a static-memory change, so tofu halts the VM; do it with -target, one node at a time, like a disk resize.

Two disks per node: a 50 GiB system disk (xvda) and a 300 GiB Longhorn data disk (xvdb). The system disk cannot be grown in place — Talos fixes the STATE partition boundaries at install, so resizing it later means replacing the VM (tofu apply -replace), cheap for a fresh node. The Longhorn disk is the opposite: raise data_disk_gib and its UserVolumeConfig expands on the next boot. The provider resizes the VDI in place but halts the VM to do it, and a plain apply would halt all three nodes at once — so grow one node per apply, tofu apply -target='module.vm["talos-01"]', and wait for Longhorn to report it healthy before the next.

Longhorn's data disk is mounted by a UserVolumeConfig named longhorn, which forces the mount to /var/mnt/longhorn; Longhorn's Helm defaultDataPath must match. The disk is selected by !system_disk rather than a device path, so Xen attach order cannot misroute it. machine.disks is deprecated from Talos 1.10 on and UserVolumeConfig is its replacement.

The kubelet needs a bind mount for that path as well, or Longhorn cannot publish volumes into workload pods:

machine:
  kubelet:
    extraMounts:
      - destination: /var/mnt/longhorn
        type: bind
        source: /var/mnt/longhorn
        options: [bind, rshared, rw]

Size the data disk above the largest volume it must hold: Longhorn schedules a replica only when the disk fits the whole volume, and a PVC the size of the disk does not fit it once filesystem overhead is taken. storageReservedPercentageForDefaultDisk: 0 is safe here only because the disk is dedicated to Longhorn.