Kubernetes

Talos Linux can be run as a pod in Kubernetes similar to running Talos in Docker. This can be used e.g. to run controlplane nodes inside an existing Kubernetes cluster.

Talos Linux running in Kubernetes is not full Talos Linux experience, as it is running in a container using the host’s kernel and network stack. Some operations like upgrades and reboots are not supported.

Prerequisites

• a running Kubernetes cluster
• a talos container image: ghcr.io/siderolabs/talos:v1.9.0

Machine Configuration

Machine configuration can be generated using Getting Started guide. Machine install disk will ge ignored, as the install image. The Talos version will be driven by the container image being used.

The required machine configuration patch to enable using container runtime DNS:

machine:
  features:
    hostDNS:
      enabled: true
      forwardKubeDNSToHost: true

Talos and Kubernetes API can be exposed using Kubernetes services or load balancers, so they can be accessed from outside the cluster.

Running Talos Pods

There might be many ways to run Talos in Kubernetes (StatefulSet, Deployment, single Pod), so we will only provide some basic guidance here.

Container Settings

env:
  - name: PLATFORM
    value: container
image: ghcr.io/siderolabs/talos:v1.9.0
ports:
  - containerPort: 50000
    name: talos-api
    protocol: TCP
  - containerPort: 6443
    name: k8s-api
    protocol: TCP
securityContext:
  privileged: true
  readOnlyRootFilesystem: true
  seccompProfile:
      type: Unconfined

Submitting Initial Machine Configuration

Initial machine configuration can be submitted using talosctl apply-config --insecure when the pod is running, or it can be submitted via an environment variable USERDATA with base64-encoded machine configuration.

Volume Mounts

Three ephemeral mounts are required for /run, /system, and /tmp directories:

volumeMounts:
  - mountPath: /run
    name: run
  - mountPath: /system
    name: system
  - mountPath: /tmp
    name: tmp

volumes:
  - emptyDir: {}
    name: run
  - emptyDir: {}
    name: system
  - emptyDir: {}
    name: tmp

Several other mountpoints are required, and they should persist across pod restarts, so one should use PersistentVolume for them:

volumeMounts:
  - mountPath: /system/state
    name: system-state
  - mountPath: /var
    name: var
  - mountPath: /etc/cni
    name: etc-cni
  - mountPath: /etc/kubernetes
    name: etc-kubernetes
  - mountPath: /usr/libexec/kubernetes
    name: usr-libexec-kubernetes