1 - Containerd

Customize Containerd Settings

The base containerd configuration expects to merge in any additional configs present in /var/cri/conf.d/*.toml.

An example of exposing metrics

Into each machine config, add the following:

    - content: |
          address = ""        
      path: /var/cri/conf.d/metrics.toml
      op: create

Create cluster like normal and see that metrics are now present on this port:

$ curl
# HELP container_blkio_io_service_bytes_recursive_bytes The blkio io service bytes recursive
# TYPE container_blkio_io_service_bytes_recursive_bytes gauge
container_blkio_io_service_bytes_recursive_bytes{container_id="0677d73196f5f4be1d408aab1c4125cf9e6c458a4bea39e590ac779709ffbe14",device="/dev/dm-0",major="253",minor="0",namespace="k8s.io",op="Async"} 0
container_blkio_io_service_bytes_recursive_bytes{container_id="0677d73196f5f4be1d408aab1c4125cf9e6c458a4bea39e590ac779709ffbe14",device="/dev/dm-0",major="253",minor="0",namespace="k8s.io",op="Discard"} 0

2 - Custom Certificate Authorities

How to supply custom certificate authorities

Appending the Certificate Authority

Put into each machine the PEM encoded certificate:

    - content: |
        -----BEGIN CERTIFICATE-----
        -----END CERTIFICATE-----        
      permissions: 0644
      path: /etc/ssl/certs/ca-certificates
      op: append

3 - Disk Encryption

Guide on using system disk encryption

It is possible to enable encryption for system disks at the OS level. As of this writing, only STATE and EPHEMERAL partitions can be encrypted. STATE contains the most sensitive node data: secrets and certs. EPHEMERAL partition may contain some sensitive workload data. Data is encrypted using LUKS2, which is provided by the Linux kernel modules and cryptsetup utility. The operating system will run additional setup steps when encryption is enabled.

If the disk encryption is enabled for the STATE partition, the system will:

  • Save STATE encryption config as JSON in the META partition.
  • Before mounting the STATE partition, load encryption configs either from the machine config or from the META partition. Note that the machine config is always preferred over the META one.
  • Before mounting the STATE partition, format and encrypt it. This occurs only if the STATE partition is empty and has no filesystem.

If the disk encryption is enabled for the EPHEMERAL partition, the system will:

  • Get the encryption config from the machine config.
  • Before mounting the EPHEMERAL partition, encrypt and format it. This occurs only if the EPHEMERAL partition is empty and has no filesystem.


Right now this encryption is disabled by default. To enable disk encryption you should modify the machine configuration with the following options:

      provider: luks2
        - nodeID: {}
          slot: 0
      provider: luks2
        - nodeID: {}
          slot: 0

Encryption Keys

Note: What the LUKS2 docs call “keys” are, in reality, a passphrase. When this passphrase is added, LUKS2 runs argon2 to create an actual key from that passphrase.

LUKS2 supports up to 32 encryption keys and it is possible to specify all of them in the machine configuration. Talos always tries to sync the keys list defined in the machine config with the actual keys defined for the LUKS2 partition. So if you update the keys list you should have at least one key that is not changed to be used for keys management.

When you define a key you should specify the key kind and the slot:

      - nodeID: {} # key kind
        slot: 1

      - static:
          passphrase: supersecret
        slot: 0

Take a note that key order does not play any role on which key slot is used. Every key must always have a slot defined.

Encryption Key Kinds

Talos supports two kinds of keys:

  • nodeID which is generated using the node UUID and the partition label (note that if the node UUID is not really random it will fail the entropy check).
  • static which you define right in the configuration.

Note: Use static keys only if your STATE partition is encrypted and only for the EPHEMERAL partition. For the STATE partition it will be stored in the META partition, which is not encrypted.

Key Rotation

It is necessary to do talosctl apply-config a couple of times to rotate keys, since there is a need to always maintain a single working key while changing the other keys around it.

So, for example, first add a new key:

      - static:
          passphrase: oldkey
        slot: 0
      - static:
          passphrase: newkey
        slot: 1


talosctl apply-config -n <node> -f config.yaml

Then remove the old key:

      - static:
          passphrase: newkey
        slot: 1


talosctl apply-config -n <node> -f config.yaml

Going from Unencrypted to Encrypted and Vice Versa

Ephemeral Partition

There is no in-place encryption support for the partitions right now, so to avoid losing any data only empty partitions can be encrypted.

As such, migration from unencrypted to encrypted needs some additional handling, especially around explicitly wiping partitions.

  • apply-config should be called with --mode=staged.
  • Partition should be wiped after apply-config, but before the reboot.

Edit your machine config and add the encryption configuration:

vim config.yaml

Apply the configuration with --mode=staged:

talosctl apply-config -f config.yaml -n <node ip> --mode=staged

Wipe the partition you’re going to encrypt:

talosctl reset --system-labels-to-wipe EPHEMERAL -n <node ip> --reboot=true

That’s it! After you run the last command, the partition will be wiped and the node will reboot. During the next boot the system will encrypt the partition.

State Partition

Calling wipe against the STATE partition will make the node lose the config, so the previous flow is not going to work.

The flow should be to first wipe the STATE partition:

talosctl reset  --system-labels-to-wipe STATE -n <node ip> --reboot=true

Node will enter into maintenance mode, then run apply-config with --insecure flag:

talosctl apply-config --insecure -n <node ip> -f config.yaml

After installation is complete the node should encrypt the STATE partition.

4 - Editing Machine Configuration

How to edit and patch Talos machine configuration, with reboot, immediately, or stage update on reboot.

Talos node state is fully defined by machine configuration. Initial configuration is delivered to the node at bootstrap time, but configuration can be updated while the node is running.

Note: Be sure that config is persisted so that configuration updates are not overwritten on reboots. Configuration persistence was enabled by default since Talos 0.5 (persist: true in machine configuration).

There are three talosctl commands which facilitate machine configuration updates:

  • talosctl apply-config to apply configuration from the file
  • talosctl edit machineconfig to launch an editor with existing node configuration, make changes and apply configuration back
  • talosctl patch machineconfig to apply automated machine configuration via JSON patch

Each of these commands can operate in one of four modes:

  • apply change in automatic mode(default): reboot if the change can’t be applied without a reboot, otherwise apply the change immediately
  • apply change with a reboot (--mode=reboot): update configuration, reboot Talos node to apply configuration change
  • apply change immediately (--mode=no-reboot flag): change is applied immediately without a reboot, fails if the change contains any fields that can not be updated without a reboot
  • apply change on next reboot (--mode=staged): change is staged to be applied after a reboot, but node is not rebooted
  • apply change in the interactive mode (--mode=interactive; only for talosctl apply-config): launches TUI based interactive installer

Note: applying change on next reboot (--mode=staged) doesn’t modify current node configuration, so next call to talosctl edit machineconfig --mode=staged will not see changes

Additionally, there is also talosctl get machineconfig, which retrieves the current node configuration API resource and contains the machine configuration in the .spec field. It can be used to modify the configuration locally before being applied to the node.

The list of config changes allowed to be applied immediately in Talos v1.1.1:

  • .debug
  • .cluster
  • .machine.time
  • .machine.certCANs
  • .machine.install (configuration is only applied during install/upgrade)
  • .machine.network
  • .machine.sysfs
  • .machine.sysctls
  • .machine.logging
  • .machine.controlplane
  • .machine.kubelet
  • .machine.pods
  • .machine.kernel
  • .machine.registries (CRI containerd plugin will not pick up the registry authentication settings without a reboot)

talosctl apply-config

This command is mostly used to submit initial machine configuration to the node (generated by talosctl gen config). It can be used to apply new configuration from the file to the running node as well, but most of the time it’s not convenient, as it doesn’t operate on the current node machine configuration.


talosctl -n <IP> apply-config -f config.yaml

Command apply-config can also be invoked as apply machineconfig:

talosctl -n <IP> apply machineconfig -f config.yaml

Applying machine configuration immediately (without a reboot):

talosctl -n IP apply machineconfig -f config.yaml --mode=no-reboot

Starting the interactive installer:

talosctl -n IP apply machineconfig --mode=interactive

Note: when a Talos node is running in the maintenance mode it’s necessary to provide --insecure (-i) flag to connect to the API and apply the config.

taloctl edit machineconfig

Command talosctl edit loads current machine configuration from the node and launches configured editor to modify the config. If config hasn’t been changed in the editor (or if updated config is empty), update is not applied.

Note: Talos uses environment variables TALOS_EDITOR, EDITOR to pick up the editor preference. If environment variables are missing, vi editor is used by default.


talosctl -n <IP> edit machineconfig

Configuration can be edited for multiple nodes if multiple IP addresses are specified:

talosctl -n <IP1>,<IP2>,... edit machineconfig

Applying machine configuration change immediately (without a reboot):

talosctl -n <IP> edit machineconfig --mode=no-reboot

talosctl patch machineconfig

Command talosctl patch works similar to talosctl edit command - it loads current machine configuration, but instead of launching configured editor it applies a set of JSON patches to the configuration and writes the result back to the node.

Example, updating kubelet version (in auto mode):

$ talosctl -n <IP> patch machineconfig -p '[{"op": "replace", "path": "/machine/kubelet/image", "value": "ghcr.io/siderolabs/kubelet:v1.24.2"}]'
patched mc at the node <IP>

Updating kube-apiserver version in immediate mode (without a reboot):

$ talosctl -n <IP> patch machineconfig --mode=no-reboot -p '[{"op": "replace", "path": "/cluster/apiServer/image", "value": "k8s.gcr.io/kube-apiserver:v1.24.2"}]'
patched mc at the node <IP>

A patch might be applied to multiple nodes when multiple IPs are specified:

talosctl -n <IP1>,<IP2>,... patch machineconfig -p '[{...}]'

Patches can also be sourced from files using @file syntax:

talosctl -n <IP> patch machineconfig -p @kubelet-patch.json -p @manifest-patch.json

It might be easier to store patches in YAML format vs. the default JSON format. Talos can detect file format automatically:

# kubelet-patch.yaml
- op: replace
  path: /machine/kubelet/image
  value: ghcr.io/siderolabs/kubelet:v1.24.2
talosctl -n <IP> patch machineconfig -p @kubelet-patch.yaml

Recovering from Node Boot Failures

If a Talos node fails to boot because of wrong configuration (for example, control plane endpoint is incorrect), configuration can be updated to fix the issue.

5 - Logging

Dealing with Talos Linux logs.

Viewing logs

Kernel messages can be retrieved with talosctl dmesg command:

$ talosctl -n dmesg kern:    info: [2021-11-10T10:09:37.662764956Z]: Command line: init_on_alloc=1 slab_nomerge pti=on consoleblank=0 nvme_core.io_timeout=4294967295 random.trust_cpu=on printk.devkmsg=on ima_template=ima-ng ima_appraise=fix ima_hash=sha512 console=ttyS0 reboot=k panic=1 talos.shutdown=halt talos.platform=metal talos.config=

Service logs can be retrieved with talosctl logs command:

$ talosctl -n services

NODE         SERVICE      STATE     HEALTH   LAST CHANGE   LAST EVENT   apid         Running   OK       19m27s ago    Health check successful   containerd   Running   OK       19m29s ago    Health check successful   cri          Running   OK       19m27s ago    Health check successful   etcd         Running   OK       19m22s ago    Health check successful   kubelet      Running   OK       19m20s ago    Health check successful   machined     Running   ?        19m30s ago    Service started as goroutine   trustd       Running   OK       19m27s ago    Health check successful   udevd        Running   OK       19m28s ago    Health check successful

$ talosctl -n logs machined [talos] task setupLogger (1/1): done, 106.109µs [talos] phase logger (1/7): done, 564.476µs

Container logs for Kubernetes pods can be retrieved with talosctl logs -k command:

$ talosctl -n containers -k
NODE         NAMESPACE   ID                                                 IMAGE                                                         PID    STATUS   k8s.io      kube-system/kube-flannel-dk6d5                     k8s.gcr.io/pause:3.5                                          1329   SANDBOX_READY   k8s.io      └─ kube-system/kube-flannel-dk6d5:install-cni      ghcr.io/siderolabs/install-cni:v0.7.0-alpha.0-1-g2bb2efc      0      CONTAINER_EXITED   k8s.io      └─ kube-system/kube-flannel-dk6d5:install-config   quay.io/coreos/flannel:v0.13.0                                0      CONTAINER_EXITED   k8s.io      └─ kube-system/kube-flannel-dk6d5:kube-flannel     quay.io/coreos/flannel:v0.13.0                                1610   CONTAINER_RUNNING   k8s.io      kube-system/kube-proxy-gfkqj                       k8s.gcr.io/pause:3.5                                          1311   SANDBOX_READY   k8s.io      └─ kube-system/kube-proxy-gfkqj:kube-proxy         k8s.gcr.io/kube-proxy:v1.24.2                                 1379   CONTAINER_RUNNING

$ talosctl -n logs -k kube-system/kube-proxy-gfkqj:kube-proxy 2021-11-30T19:13:20.567825192Z stderr F I1130 19:13:20.567737       1 server_others.go:138] "Detected node IP" address="" 2021-11-30T19:13:20.599684397Z stderr F I1130 19:13:20.599613       1 server_others.go:206] "Using iptables Proxier"

Sending logs

Service logs

You can enable logs sendings in machine configuration:

      - endpoint: "udp://"
        format: "json_lines"
      - endpoint: "tcp://host:5044/"
        format: "json_lines"

Several destinations can be specified. Supported protocols are UDP and TCP. The only currently supported format is json_lines:

  "msg": "[talos] apply config request: immediate true, on reboot false",
  "talos-level": "info",
  "talos-service": "machined",
  "talos-time": "2021-11-10T10:48:49.294858021Z"

Messages are newline-separated when sent over TCP. Over UDP messages are sent with one message per packet. msg, talos-level, talos-service, and talos-time fields are always present; there may be additional fields.

Kernel logs

Kernel log delivery can be enabled with the talos.logging.kernel kernel command line argument, which can be specified in the .machine.installer.extraKernelArgs:

      - talos.logging.kernel=tcp://host:5044/

Kernel log destination is specified in the same way as service log endpoint. The only supported format is json_lines.

Sample message:

  "clock":6252819, // time relative to the kernel boot time
  "msg":"[talos] task startAllServices (1/1): waiting for 6 services\n",
  "talos-level":"warn", // Talos-translated `priority` into common logging level
  "talos-time":"2021-11-26T16:53:21.3258698Z" // Talos-translated `clock` using current time

extraKernelArgs in the machine configuration are only applied on Talos upgrades, not just by applying the config. (Upgrading to the same version is fine).

Filebeat example

To forward logs to other Log collection services, one way to do this is sending them to a Filebeat running in the cluster itself (in the host network), which takes care of forwarding it to other endpoints (and the necessary transformations).

If Elastic Cloud on Kubernetes is being used, the following Beat (custom resource) configuration might be helpful:

apiVersion: beat.k8s.elastic.co/v1beta1
kind: Beat
  name: talos
  type: filebeat
  version: 7.15.1
    name: talos
      - type: "udp"
        host: ""
          - decode_json_fields:
              fields: ["message"]
              target: ""
          - timestamp:
              field: "talos-time"
                - "2006-01-02T15:04:05.999999999Z07:00"
          - drop_fields:
              fields: ["message", "talos-time"]
          - rename:
                - from: "msg"
                  to: "message"

        maxUnavailable: 100%
        dnsPolicy: ClusterFirstWithHostNet
        hostNetwork: true
          runAsUser: 0
          - name: filebeat
              - protocol: UDP
                containerPort: 12345
                hostPort: 12345

The input configuration ensures that messages and timestamps are extracted properly. Refer to the Filebeat documentation on how to forward logs to other outputs.

Also note the hostNetwork: true in the daemonSet configuration.

This ensures filebeat uses the host network, and listens on (UDP) on every machine, which can then be specified as a logging endpoint in the machine configuration.

Fluent-bit example

First, we’ll create a value file for the fluentd-bit Helm chart.

# fluentd-bit.yaml

  fluentbit.io/exclude: 'true'

  - port: 12345
    containerPort: 12345
    protocol: TCP
    name: talos

  service: |
      Flush         5
      Daemon        Off
      Log_Level     warn
      Parsers_File  custom_parsers.conf    

  inputs: |
      Name          tcp
      Port          12345
      Format        json
      Tag           talos.*

      Name          tail
      Alias         kubernetes
      Path          /var/log/containers/*.log
      Parser        containerd
      Tag           kubernetes.*

      Name          tail
      Alias         audit
      Path          /var/log/audit/kube/*.log
      Parser        audit
      Tag           audit.*    

  filters: |
      Name                kubernetes
      Alias               kubernetes
      Match               kubernetes.*
      Kube_Tag_Prefix     kubernetes.var.log.containers.
      Use_Kubelet         Off
      Merge_Log           On
      Merge_Log_Trim      On
      Keep_Log            Off
      K8S-Logging.Parser  Off
      K8S-Logging.Exclude On
      Annotations         Off
      Labels              On

      Name          modify
      Match         kubernetes.*
      Add           source kubernetes
      Remove        logtag    

  customParsers: |
      Name          audit
      Format        json
      Time_Key      requestReceivedTimestamp
      Time_Format   %Y-%m-%dT%H:%M:%S.%L%z

      Name          containerd
      Format        regex
      Regex         ^(?<time>[^ ]+) (?<stream>stdout|stderr) (?<logtag>[^ ]*) (?<log>.*)$
      Time_Key      time
      Time_Format   %Y-%m-%dT%H:%M:%S.%L%z    

  outputs: |
      Name    stdout
      Alias   stdout
      Match   *
      Format  json_lines    

  # If you wish to ship directly to Loki from Fluentbit,
  # Uncomment the following output, updating the Host with your Loki DNS/IP info as necessary.
  # [OUTPUT]
  # Name loki
  # Match *
  # Host loki.loki.svc
  # Port 3100
  # Labels job=fluentbit
  # Auto_Kubernetes_Labels on

  - name: varlog
      path: /var/log

  - name: varlog
    mountPath: /var/log

  - operator: Exists
    effect: NoSchedule

Next, we will add the helm repo for FluentBit, and deploy it to the cluster.

helm repo add fluent https://fluent.github.io/helm-charts
helm upgrade -i --namespace=kube-system -f fluentd-bit.yaml fluent-bit fluent/fluent-bit

Now we need to find the service IP.

$ kubectl -n kube-system get svc -l app.kubernetes.io/name=fluent-bit

NAME         TYPE        CLUSTER-IP     EXTERNAL-IP   PORT(S)             AGE
fluent-bit   ClusterIP   <none>        2020/TCP,5170/TCP   108m

Finally, we will change talos log destination with the command talosctl edit mc.

      - endpoint: "tcp://"
        format: "json_lines"

This example configuration was well tested with Cilium CNI, and it should work with iptables/ipvs based CNI plugins too.

Vector example

Vector is a lightweight observability pipeline ideal for a Kubernetes environment. It can ingest (source) logs from multiple sources, perform remapping on the logs (transform), and forward the resulting pipeline to multiple destinations (sinks). As it is an end to end platform, it can be run as a single-deployment ‘aggregator’ as well as a replicaSet of ‘Agents’ that run on each node.

As Talos can be set as above to send logs to a destination, we can run Vector as an Aggregator, and forward both kernel and service to a UDP socket in-cluster.

Below is an excerpt of a source/sink setup for Talos, with a ‘sink’ destination of an in-cluster Grafana Loki log aggregation service. As Loki can create labels from the log input, we have set up the Loki sink to create labels based on the host IP, service and facility of the inbound logs.

Note that a method of exposing the Vector service will be required which may vary depending on your setup - a LoadBalancer is a good option.

role: "Stateless-Aggregator"

# Sources
    type: socket
    mode: udp
    max_length: 102400
      codec: json
    host_key: __host

    type: socket
    mode: udp
    max_length: 102400
      codec: json
    host_key: __host

# Sinks
    type: loki
      - talos_kernel_logs_xform
    endpoint: http://loki.system-monitoring:3100
      codec: json
        - __host
      max_bytes: 1048576
    out_of_order_action: rewrite_timestamp
      hostname: >-
                {{`{{ __host }}`}}
      facility: >-
                {{`{{ facility }}`}}

    type: loki
      - talos_service_logs_xform
    endpoint: http://loki.system-monitoring:3100
      codec: json
        - __host
      max_bytes: 400000
    out_of_order_action: rewrite_timestamp
      hostname: >-
                {{`{{ __host }}`}}
      service: >-
                {{`{{ "talos-service" }}`}}

6 - Managing PKI

How to manage Public Key Infrastructure

Generating an Administrator Key Pair

In order to create a key pair, you will need the root CA.

Save the CA public key, and CA private key as ca.crt, and ca.key respectively. Now, run the following commands to generate a certificate:

talosctl gen key --name admin
talosctl gen csr --key admin.key --ip
talosctl gen crt --ca ca --csr admin.csr --name admin

Now, base64 encode admin.crt, and admin.key:

cat admin.crt | base64
cat admin.key | base64

You can now set the crt and key fields in the talosconfig to the base64 encoded strings.

Renewing an Expired Administrator Certificate

In order to renew the certificate, you will need the root CA, and the admin private key. The base64 encoded key can be found in any one of the control plane node’s configuration file. Where it is exactly will depend on the specific version of the configuration file you are using.

Save the CA public key, CA private key, and admin private key as ca.crt, ca.key, and admin.key respectively. Now, run the following commands to generate a certificate:

talosctl gen csr --key admin.key --ip
talosctl gen crt --ca ca --csr admin.csr --name admin

You should see admin.crt in your current directory. Now, base64 encode admin.crt:

cat admin.crt | base64

You can now set the certificate in the talosconfig to the base64 encoded string.


In this guide we’ll follow the procedure to support NVIDIA GPU on Talos.

Enabling NVIDIA GPU support on Talos is bound by NVIDIA EULA Talos GPU support is an alpha feature.

These are the steps to enabling NVIDIA support in Talos.

  • Talos pre-installed on a node with NVIDIA GPU installed.
  • Building a custom Talos installer image with NVIDIA modules
  • Building NVIDIA container toolkit system extension which allows to register a custom runtime with containerd
  • Upgrading Talos with the custom installer and enabling NVIDIA modules and the system extension

Both these components require that the user build and maintain their own Talos installer image and the NVIDIA container toolkit Talos System Extension.


This guide assumes the user has access to a container registry with push permissions, docker installed on the build machine and the Talos host has pull access to the container registry.

Set the local registry and username environment variables:

export USERNAME=<username>
export REGISTRY=<registry>

For eg:

export USERNAME=talos-user
export REGISTRY=ghcr.io

The examples below will use the sample variables set above. Modify accordingly for your environment.

Building the installer image

Start by cloning the pkgs repository.

Now run the following command to build and push custom Talos kernel image and the NVIDIA image with the NVIDIA kernel modules signed by the kernel built along with it.

make kernel nonfree-kmod-nvidia PLATFORM=linux/amd64 PUSH=true

Replace the platform with linux/arm64 if building for ARM64

Now we need to create a custom Talos installer image.

Start by creating a Dockerfile with the following content:

FROM scratch as customization
COPY --from=ghcr.io/talos-user/nonfree-kmod-nvidia:v1.1.1-nvidia /lib/modules /lib/modules

FROM ghcr.io/siderolabs/installer:v1.1.1
COPY --from=ghcr.io/talos-user/kernel:v1.1.1-nvidia /boot/vmlinuz /usr/install/${TARGETARCH}/vmlinuz

Now build the image and push it to the registry.

DOCKER_BUILDKIT=0 docker build --squash --build-arg RM="/lib/modules" -t ghcr.io/talos-user/installer:v1.1.1-nvidia .
docker push ghcr.io/talos-user/installer:v1.1.1-nvidia

Note: buildkit has a bug #816, to disable it use DOCKER_BUILDKIT=0

Building the system extension

Start by cloning the extensions repository.

Now run the following command to build and push the system extension.

make nvidia-container-toolkit PLATFORM=linux/amd64 PUSH=true TAG=510.60.02-v1.9.0

Replace the platform with linux/arm64 if building for ARM64

Upgrading Talos and enabling the NVIDIA modules and the system extension

Make sure to use talosctl version v1.1.1 or later

First create a patch yaml gpu-worker-patch.yaml to update the machine config similar to below:

- op: add
  path: /machine/install/extensions
    - image: ghcr.io/talos-user/nvidia-container-toolkit:510.60.02-v1.9.0
- op: add
  path: /machine/kernel
      - name: nvidia
      - name: nvidia_uvm
      - name: nvidia_drm
      - name: nvidia_modeset
- op: add
  path: /machine/sysctls
    net.core.bpf_jit_harden: 1

Now apply the patch to all Talos nodes in the cluster having NVIDIA GPU’s installed:

talosctl patch mc --patch @gpu-worker-patch.yaml

Now we can proceed to upgrading Talos with the installer built previously:

talosctl upgrade --image=ghcr.io/talos-user/installer:v1.1.1-nvidia

Once the node reboots, the NVIDIA modules should be loaded and the system extension should be installed.

This can be confirmed by running:

talosctl read /proc/modules

which should produce an output similar to below:

nvidia_uvm 1146880 - - Live 0xffffffffc2733000 (PO)
nvidia_drm 69632 - - Live 0xffffffffc2721000 (PO)
nvidia_modeset 1142784 - - Live 0xffffffffc25ea000 (PO)
nvidia 39047168 - - Live 0xffffffffc00ac000 (PO)
talosctl get extensions

which should produce an output similar to below:

NODE           NAMESPACE   TYPE              ID                                                                 VERSION   NAME                       VERSION   runtime     ExtensionStatus   000.ghcr.io-frezbo-nvidia-container-toolkit-510.60.02-v1.9.0       1         nvidia-container-toolkit   510.60.02-v1.9.0
talosctl read /proc/driver/nvidia/version

which should produce an output similar to below:

NVRM version: NVIDIA UNIX x86_64 Kernel Module  510.60.02  Wed Mar 16 11:24:05 UTC 2022
GCC version:  gcc version 11.2.0 (GCC)

Deploying NVIDIA device plugin

First we need to create the RuntimeClass

Apply the following manifest to create a runtime class that uses the extension:

apiVersion: node.k8s.io/v1
kind: RuntimeClass
  name: nvidia
handler: nvidia

Install the NVIDIA device plugin:

helm repo add nvdp https://nvidia.github.io/k8s-device-plugin
helm repo update
helm install nvidia-device-plugin nvdp/nvidia-device-plugin --version=0.11.0 --set=runtimeClassName=nvidia

Apply the following manifest to run CUDA pod via nvidia runtime:

cat <<EOF | kubectl apply -f -
apiVersion: v1
kind: Pod
  name: gpu-operator-test
  restartPolicy: OnFailure
  runtimeClassName: nvidia
  - name: cuda-vector-add
    image: "nvidia/samples:vectoradd-cuda11.6.0"
         nvidia.com/gpu: 1

The status can be viewed by running:

kubectl get pods

which should produce an output similar to below:

NAME                READY   STATUS      RESTARTS   AGE
gpu-operator-test   0/1     Completed   0          13s
kubectl logs gpu-operator-test

which should produce an output similar to below:

[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory

8 - Pull Through Image Cache

How to set up local transparent container images caches.

In this guide we will create a set of local caching Docker registry proxies to minimize local cluster startup time.

When running Talos locally, pulling images from Docker registries might take a significant amount of time. We spin up local caching pass-through registries to cache images and configure a local Talos cluster to use those proxies. A similar approach might be used to run Talos in production in air-gapped environments. It can be also used to verify that all the images are available in local registries.

Video Walkthrough

To see a live demo of this writeup, see the video below:


The follow are requirements for creating the set of caching proxies:

  • Docker 18.03 or greater
  • Local cluster requirements for either docker or QEMU.

Launch the Caching Docker Registry Proxies

Talos pulls from docker.io, k8s.gcr.io, quay.io, gcr.io, and ghcr.io by default. If your configuration is different, you might need to modify the commands below:

docker run -d -p 5000:5000 \
    -e REGISTRY_PROXY_REMOTEURL=https://registry-1.docker.io \
    --restart always \
    --name registry-docker.io registry:2

docker run -d -p 5001:5000 \
    -e REGISTRY_PROXY_REMOTEURL=https://k8s.gcr.io \
    --restart always \
    --name registry-k8s.gcr.io registry:2

docker run -d -p 5002:5000 \
    -e REGISTRY_PROXY_REMOTEURL=https://quay.io \
    --restart always \
    --name registry-quay.io registry:2.5

docker run -d -p 5003:5000 \
    -e REGISTRY_PROXY_REMOTEURL=https://gcr.io \
    --restart always \
    --name registry-gcr.io registry:2

docker run -d -p 5004:5000 \
    -e REGISTRY_PROXY_REMOTEURL=https://ghcr.io \
    --restart always \
    --name registry-ghcr.io registry:2

Note: Proxies are started as docker containers, and they’re automatically configured to start with Docker daemon. Please note that quay.io proxy doesn’t support recent Docker image schema, so we run older registry image version (2.5).

As a registry container can only handle a single upstream Docker registry, we launch a container per upstream, each on its own host port (5000, 5001, 5002, 5003 and 5004).

Using Caching Registries with QEMU Local Cluster

With a QEMU local cluster, a bridge interface is created on the host. As registry containers expose their ports on the host, we can use bridge IP to direct proxy requests.

sudo talosctl cluster create --provisioner qemu \
    --registry-mirror docker.io= \
    --registry-mirror k8s.gcr.io= \
    --registry-mirror quay.io= \
    --registry-mirror gcr.io= \
    --registry-mirror ghcr.io=

The Talos local cluster should now start pulling via caching registries. This can be verified via registry logs, e.g. docker logs -f registry-docker.io. The first time cluster boots, images are pulled and cached, so next cluster boot should be much faster.

Note: is a bridge IP with default network (, if using custom --cidr, value should be adjusted accordingly.

Using Caching Registries with docker Local Cluster

With a docker local cluster we can use docker bridge IP, default value for that IP is On Linux, the docker bridge address can be inspected with ip addr show docker0.

talosctl cluster create --provisioner docker \
    --registry-mirror docker.io= \
    --registry-mirror k8s.gcr.io= \
    --registry-mirror quay.io= \
    --registry-mirror gcr.io= \
    --registry-mirror ghcr.io=

Cleaning Up

To cleanup, run:

docker rm -f registry-docker.io
docker rm -f registry-k8s.gcr.io
docker rm -f registry-quay.io
docker rm -f registry-gcr.io
docker rm -f registry-ghcr.io

Note: Removing docker registry containers also removes the image cache. So if you plan to use caching registries, keep the containers running.

9 - Role-based access control (RBAC)

Set up RBAC on the Talos Linux API.

Talos v0.11 introduced initial support for role-based access control (RBAC). This guide will explain what that is and how to enable it without losing access to the cluster.

RBAC in Talos

Talos uses certificates to authorize users. The certificate subject’s organization field is used to encode user roles. There is a set of predefined roles that allow access to different API methods:

  • os:admin grants access to all methods;
  • os:reader grants access to “safe” methods (for example, that includes the ability to list files, but does not include the ability to read files content);
  • os:etcd:backup grants access to /machine.MachineService/EtcdSnapshot method.

Roles in the current talosconfig can be checked with the following command:

$ talosctl config info

Roles:               os:admin

RBAC is enabled by default in new clusters created with talosctl v0.11+ and disabled otherwise.

Enabling RBAC

First, both the Talos cluster and talosctl tool should be upgraded. Then the talosctl config new command should be used to generate a new client configuration with the os:admin role. Additional configurations and certificates for different roles can be generated by passing --roles flag:

talosctl config new --roles=os:reader reader

That command will create a new client configuration file reader with a new certificate with os:reader role.

After that, RBAC should be enabled in the machine configuration:

    rbac: true

10 - System Extensions

Customizing the Talos Linux immutable root file system.

System extensions allow extending the Talos root filesystem, which enables a variety of features, such as including custom container runtimes, loading additional firmware, etc.

System extensions are only activated during the installation or upgrade of Talos Linux. With system extensions installed, the Talos root filesystem is still immutable and read-only.


System extensions are configured in the .machine.install section:

      - image: ghcr.io/siderolabs/gvisor:33f613e

During the initial install (e.g. when PXE booting or booting from an ISO), Talos will pull down container images for system extensions, validate them, and include them into the Talos initramfs image. System extensions will be activated on boot and overlaid on top of the Talos root filesystem.

In order to update the system extensions for a running instance, update .machine.install.extensions and upgrade Talos. (Note: upgrading to the same version of Talos is fine).

Building a Talos Image with System Extensions

System extensions can be installed into the Talos disk image (e.g. AWS AMI or VMWare OVF) by running the following command to generate the image from the Talos source tree:

make image-metal IMAGER_SYSTEM_EXTENSIONS="ghcr.io/siderolabs/amd-ucode:20220411 ghcr.io/siderolabs/gvisor:20220405.0-v1.0.0-10-g82b41ad"

Authoring System Extensions

A Talos system extension is a container image with the specific folder structure. System extensions can be built and managed using any tool that produces container images, e.g. docker build.

Sidero Labs maintains a repository of system extensions.

Resource Definitions

Use talosctl get extensions to get a list of system extensions:

$ talosctl get extensions
NODE         NAMESPACE   TYPE              ID                                              VERSION   NAME          VERSION   runtime     ExtensionStatus   000.ghcr.io-talos-systems-gvisor-54b831d        1         gvisor        20220117.0-v1.0.0   runtime     ExtensionStatus   001.ghcr.io-talos-systems-intel-ucode-54b831d   1         intel-ucode   microcode-20210608-v1.0.0

Use YAML or JSON format to see additional details about the extension:

$ talosctl -n get extensions 001.ghcr.io-talos-systems-intel-ucode-54b831d -o yaml
    namespace: runtime
    type: ExtensionStatuses.runtime.talos.dev
    id: 001.ghcr.io-talos-systems-intel-ucode-54b831d
    version: 1
    owner: runtime.ExtensionStatusController
    phase: running
    created: 2022-02-10T18:25:04Z
    updated: 2022-02-10T18:25:04Z
    image: 001.ghcr.io-talos-systems-intel-ucode-54b831d.sqsh
        name: intel-ucode
        version: microcode-20210608-v1.0.0
        author: Spencer Smith
        description: |
            This system extension provides Intel microcode binaries.
                version: '>= v1.0.0'

Example: gVisor

See readme of the gVisor extension.