Cluster Nodes

Node Configuration

All nodes are QEMU VMs on a Proxmox host, defined in core/platform/talos.go.

Control Plane Nodes

Control plane nodes carry a floating VIP (192.168.1.210) on their primary interface. This VIP is configured as a Talos machine patch so the API server endpoint remains stable regardless of which node holds it.

Worker Nodes

Workers use 200 GiB disks to provide ~90–95 GiB of usable Longhorn storage per node. The GPU worker (k8s-worker4) has a 250 GiB disk for extra AI model volume space.

Talos Version

Talos Linux — cluster name talos-cluster; version in the Software Inventory.

Talos Images

Two Talos image variants are used:

Base image (schematic 88d1f7a5c4f1d3aba7df787c448c1d3d008ed29cfb34af53fa0df4336a56040b):

• Extensions: iscsi-tools, util-linux-tools, qemu-guest-agent
• Used by: control plane nodes + k8s-worker1–3

GPU image (schematic 901b9afcf2f7eda57991690fc5ca00414740cc4ee4ad516109bcc58beff1b829):

• Extensions: all base extensions + nvidia-container-toolkit-production, nvidia-open-gpu-kernel-modules-production
• Used by: k8s-worker4

Worker Node Kernel Modules

All workers load the following kernel modules via Talos machine config patches:

nbd, iscsi_tcp, iscsi_generic, configfs

The GPU worker (k8s-worker4) additionally loads:

nvidia, nvidia_uvm, nvidia_drm, nvidia_modeset

Network Configuration

Nodes take their address by DHCP, not static machine config. The addresses in this document are DHCP reservations on the router — Talos itself just asks:

machine:
  network:
    interfaces:
      - deviceSelector:
          physical: true
        dhcp: true

Changing a node's IP therefore means changing the reservation, not the Talos config. Control planes additionally get the shared VIP on the same interface.

Longhorn Node Labels

All worker nodes carry the label node.longhorn.io/create-default-disk: config, applied via the Talos worker machine patch. This tells Longhorn to provision its default disk on every worker.