InfrastructureDesign phase

Talos Kubernetes Platform

Three Talos clusters across a 7-host vSphere estate. Design freeze, IP plan, and failure tests written before any VM exists.

A design package (failure model, IP plan, placement map, failure-test catalogue, secrets policy) for three immutable Kubernetes clusters.

Role
Solo: design
Timeframe
May 2026
repo created 2026-05-23
Team
1
Problem
Talos has no shell, so configuration has to be right before a machine boots. That makes the plan the actual deliverable.
Approach
Write the architecture, the IP plan and the failure tests first, and keep generated secrets out of git by policy.
Result
Design only. The three environments I actually run today are RKE2 (see the three-environment platform page); the Talos design informed their IP plan and placement rules.
3
Clusters
mgmt / east / west
3+3
Topology
control plane + workers each
7
ESXi hosts
cluster Compute-01
110/120/130
VLANs
one per cluster

Every figure above is measured or read from the running system. Where a figure has a source line, that is how and when it was read.

Technology Stack

Talos LinuxOSKubernetesOrchestrationvSphere / ESXiVirtualizationOpenTofuIaCgovcTooling

What it is

Three immutable Talos Kubernetes clusters (a management cluster and two workload clusters) planned across my vSphere estate, each on its own VLAN with its own API VIP.

This project is in its design phase and the page says so. What exists today is the part most homelab writeups skip: an architecture doc with an explicit failure model and stated non-goals, a full IP plan covering VLANs, pod and service CIDRs and static addresses, a host-to-VM placement map with DRS rules, a bootstrap runbook, a failure-test catalogue, and a secrets policy that keeps generated Talos configs and secrets out of git.

Why design-first

Talos has no SSH and no shell. You cannot fix a node by logging into it, which means the config has to be right before the machine boots. That property turns 'write the IP plan first' from good hygiene into a hard requirement, and it is the main reason I picked Talos for this.

Deciding the failure model on paper (what happens when a host dies, when a VLAN drops, when etcd loses quorum) is also considerably cheaper than discovering it with 18 VMs already running.