Troubleshooting Control Plane

This guide is written as series of topics and detailed answers for each topic. It starts with basics of control plane and goes into Talos specifics.

This document mostly applies only to Talos 0.9 control plane based on static pods. If Talos was upgraded from version 0.8, it might be still running self-hosted control plane, current status can be checked with the command talosctl get bootstrapstatus:

$ talosctl -n <IP> get bs
NODE         NAMESPACE   TYPE              ID              VERSION   SELF HOSTED   runtime     BootstrapStatus   control-plane   1         false

In this guide we assume that Talos client config is available and Talos API access is available. Kubernetes client configuration can be pulled from control plane nodes with talosctl -n <IP> kubeconfig (this command works before Kubernetes is fully booted).

What is a control plane node?

Talos nodes which have .machine.type of init and controlplane are control plane nodes.

The only difference between init and controlplane nodes is that init node automatically bootstraps a single-node etcd cluster on a first boot if the etcd data directory is empty. A node with type init can be replaced with a controlplane node which is triggered to run etcd bootstrap with talosctl --nodes <IP> bootstrap command.

Use of init type nodes is discouraged, as it might lead to split-brain scenario if one node in existing cluster is reinstalled while config type is still init.

It is critical to make sure only one control plane runs in bootstrap mode (either with node type init or via bootstrap API/talosctl bootstrap), as having more than node in bootstrap mode leads to split-brain scenario (multiple etcd clusters are built instead of a single cluster).

What is special about control plane node?

Control plane nodes in Talos run etcd which provides data store for Kubernetes and Kubernetes control plane components (kube-apiserver, kube-controller-manager and kube-scheduler).

Control plane nodes are tainted by default to prevent workloads from being scheduled to control plane nodes.

How many control plane nodes should be deployed?

With a single control plane node, cluster is not HA: if that single node experiences hardware failure, cluster control plane is broken and can't be recovered. Single control plane node clusters are still used as test clusters and in edge deployments, but it should be noted that this setup is not HA.

Number of control plane should be odd (1, 3, 5, ...), as with even number of nodes, etcd quorum doesn't tolerate failures correctly: e.g. with 2 control plane nodes quorum is 2, so failure of any node breaks quorum, so this setup is almost equivalent to single control plane node cluster.

With three control plane nodes cluster can tolerate a failure of any single control plane node. With five control plane nodes cluster can tolerate failure of any two control plane nodes.

What is control plane endpoint?

Kubernetes requires having a control plane endpoint which points to any healthy API server running on a control plane node. Control plane endpoint is specified as URL like https://endpoint:6443/. At any point in time, even during failures control plane endpoint should point to a healthy API server instance. As kube-apiserver runs with host network, control plane endpoint should point to one of the control plane node IPs: node1:6443, node2:6443, ...

For single control plane node clusters, control plane endpoint might be https://IP:6443/ or https://DNS:6443/, where IP is the IP of the control plane node and DNS points to IP. DNS form of the endpoint allows to change the IP address of the control plane if that IP changes over time.

For HA clusters, control plane can be implemented as:

  • TCP L7 loadbalancer with active health checks against port 6443
  • round-robin DNS with active health checks against port 6443
  • BGP anycast IP with health checks
  • virtual shared L2 IP

It is critical that control plane endpoint works correctly during cluster bootstrap phase, as nodes discover each other using control plane endpoint.

kubelet is not running on control plane node

Service kubelet should be running on control plane node as soon as networking is configured:

$ talosctl -n <IP> service kubelet
ID       kubelet
STATE    Running
EVENTS   [Running]: Health check successful (2m54s ago)
         [Running]: Health check failed: Get "": dial tcp connect: connection refused (3m4s ago)
         [Running]: Started task kubelet (PID 2334) for container kubelet (3m6s ago)
         [Preparing]: Creating service runner (3m6s ago)
         [Preparing]: Running pre state (3m15s ago)
         [Waiting]: Waiting for service "timed" to be "up" (3m15s ago)
         [Waiting]: Waiting for service "cri" to be "up", service "timed" to be "up" (3m16s ago)
         [Waiting]: Waiting for service "cri" to be "up", service "networkd" to be "up", service "timed" to be "up" (3m18s ago)

If kubelet is not running, it might be caused by wrong configuration, check kubelet logs with talosctl logs:

$ talosctl -n <IP> logs kubelet I0305 20:45:07.756948    2334 controller.go:101] kubelet config controller: starting controller I0305 20:45:07.756995    2334 controller.go:267] kubelet config controller: ensuring filesystem is set up correctly I0305 20:45:07.757000    2334 fsstore.go:59] kubelet config controller: initializing config checkpoints directory "/etc/kubernetes/kubelet/store"

etcd is not running on bootstrap node

etcd should be running on bootstrap node immediately (bootstrap node is either init node or controlplane node after talosctl bootstrap command was issued). When node boots for the first time, etcd data directory /var/lib/etcd directory is empty and Talos launches etcd in a mode to build the initial cluster of a single node. At this time /var/lib/etcd directory becomes non-empty and etcd runs as usual.

If etcd is not running, check service etcd state:

$ talosctl -n <IP> service etcd
ID       etcd
STATE    Running
EVENTS   [Running]: Health check successful (3m21s ago)
         [Running]: Started task etcd (PID 2343) for container etcd (3m26s ago)
         [Preparing]: Creating service runner (3m26s ago)
         [Preparing]: Running pre state (3m26s ago)
         [Waiting]: Waiting for service "cri" to be "up", service "networkd" to be "up", service "timed" to be "up" (3m26s ago)

If service is stuck in Preparing state for bootstrap node, it might be related to slow network - at this stage Talos pulls etcd image from the container registry.

If etcd service is crashing and restarting, check service logs with talosctl -n <IP> logs etcd. Most common reasons for crashes are:

  • wrong arguments passed via extraArgs in the configuration;
  • booting Talos on non-empty disk with previous Talos installation, /var/lib/etcd contains data from old cluster.

etcd is not running on non-bootstrap control plane node

Service etcd on non-bootstrap control plane node waits for Kubernetes to boot successfully on bootstrap node to find other peers to build a cluster. As soon as bootstrap node boots Kubernetes control plane components, and kubectl get endpoints returns IP of bootstrap control plane node, other control plane nodes will start joining the cluster followed by Kubernetes control plane components on each control plane node.

Kubernetes static pod definitions are not generated

Talos should write down static pod definitions for the Kubernetes control plane:

$ talosctl -n <IP> ls /etc/kubernetes/manifests
NODE         NAME   .   talos-kube-apiserver.yaml   talos-kube-controller-manager.yaml   talos-kube-scheduler.yaml

If static pod definitions are not rendered, check etcd and kubelet service health (see above), and controller runtime logs (talosctl logs controller-runtime).

Talos prints error an error on the server ("") has prevented the request from succeeding

This is expected during initial cluster bootstrap and sometimes after a reboot:

[   70.093289] [talos] task labelNodeAsMaster (1/1): starting
[   80.094038] [talos] retrying error: an error on the server ("") has prevented the request from succeeding (get nodes talos-default-master-1)

Initially kube-apiserver component is not running yet, and it takes some time before it becomes fully up during bootstrap (image should be pulled from the Internet, etc.) Once control plane endpoint is up Talos should proceed.

If Talos doesn't proceed further, it might be a configuration issue.

In any case, status of control plane components can be checked with talosctl containers -k:

$ talosctl -n <IP> containers --kubernetes
NODE         NAMESPACE   ID                                                                                      IMAGE                                        PID    STATUS      kube-system/kube-apiserver-talos-default-master-1                                                      2539   SANDBOX_READY      └─ kube-system/kube-apiserver-talos-default-master-1:kube-apiserver                       2572   CONTAINER_RUNNING

If kube-apiserver shows as CONTAINER_EXITED, it might have exited due to configuration error. Logs can be checked with taloctl logs --kubernetes (or with -k as a shorthand):

$ talosctl -n <IP> logs -k kube-system/kube-apiserver-talos-default-master-1:kube-apiserver 2021-03-05T20:46:13.133902064Z stderr F 2021/03/05 20:46:13 Running command: 2021-03-05T20:46:13.133933824Z stderr F Command env: (log-file=, also-stdout=false, redirect-stderr=true) 2021-03-05T20:46:13.133938524Z stderr F Run from directory: 2021-03-05T20:46:13.13394154Z stderr F Executable path: /usr/local/bin/kube-apiserver

Talos prints error nodes "talos-default-master-1" not found

This error means that kube-apiserver is up, and control plane endpoint is healthy, but kubelet hasn't got its client certificate yet and wasn't able to register itself.

For the kubelet to get its client certificate, following conditions should apply:

  • control plane endpoint is healthy (kube-apiserver is running)
  • bootstrap manifests got successfully deployed (for CSR auto-approval)
  • kube-controller-manager is running

CSR state can be checked with kubectl get csr:

$ kubectl get csr
NAME        AGE   SIGNERNAME                                    REQUESTOR                 CONDITION
csr-jcn9j   14m   system:bootstrap:q9pyzr   Approved,Issued
csr-p6b9q   14m   system:bootstrap:q9pyzr   Approved,Issued
csr-sw6rm   14m   system:bootstrap:q9pyzr   Approved,Issued
csr-vlghg   14m   system:bootstrap:q9pyzr   Approved,Issued

Talos prints error node not ready

Node in Kubernetes is marked as Ready once CNI is up. It takes a minute or two for the CNI images to be pulled and for the CNI to start. If the node is stuck in this state for too long, check CNI pods and logs with kubectl, usually CNI resources are created in kube-system namespace. For example, for Talos default Flannel CNI:

$ kubectl -n kube-system get pods
NAME                                             READY   STATUS    RESTARTS   AGE
kube-flannel-25drx                               1/1     Running   0          23m
kube-flannel-8lmb6                               1/1     Running   0          23m
kube-flannel-gl7nx                               1/1     Running   0          23m
kube-flannel-jknt9                               1/1     Running   0          23m

Talos prints error x509: certificate signed by unknown authority

Full error might look like:

x509: certificate signed by unknown authority (possiby because of crypto/rsa: verification error" while trying to verify candidate authority certificate "kubernetes"

Commonly, the control plane endpoint points to a different cluster, as the client certificate generated by Talos doesn't match CA of the cluster at control plane endpoint.

etcd is running on bootstrap node, but stuck in pre state on non-bootstrap nodes

Please see question etcd is not running on non-bootstrap control plane node.

Checking kube-controller-manager and kube-scheduler

If control plane endpoint is up, status of the pods can be performed with kubectl:

$ kubectl get pods -n kube-system -l k8s-app=kube-controller-manager
NAME                                             READY   STATUS    RESTARTS   AGE
kube-controller-manager-talos-default-master-1   1/1     Running   0          28m
kube-controller-manager-talos-default-master-2   1/1     Running   0          28m
kube-controller-manager-talos-default-master-3   1/1     Running   0          28m

If control plane endpoint is not up yet, container status can be queried with talosctl containers --kubernetes:

$ talosctl -n <IP> c -k
NODE         NAMESPACE   ID                                                                                      IMAGE                                        PID    STATUS
...      kube-system/kube-controller-manager-talos-default-master-1                                             2547   SANDBOX_READY      └─ kube-system/kube-controller-manager-talos-default-master-1:kube-controller-manager   2580   CONTAINER_RUNNING      kube-system/kube-scheduler-talos-default-master-1                                                      2638   SANDBOX_READY      └─ kube-system/kube-scheduler-talos-default-master-1:kube-scheduler                       2670   CONTAINER_RUNNING

If some of the containers are not running, it could be that image is still being pulled. Otherwise process might crashing, in that case logs can be checked with talosctl logs --kubernetes <containerID>:

$ talosctl -n <IP> logs -k kube-system/kube-controller-manager-talos-default-master-1:kube-controller-manager 2021-03-09T13:59:34.291667526Z stderr F 2021/03/09 13:59:34 Running command: 2021-03-09T13:59:34.291702262Z stderr F Command env: (log-file=, also-stdout=false, redirect-stderr=true) 2021-03-09T13:59:34.291707121Z stderr F Run from directory: 2021-03-09T13:59:34.291710908Z stderr F Executable path: /usr/local/bin/kube-controller-manager 2021-03-09T13:59:34.291719163Z stderr F Args (comma-delimited): /usr/local/bin/kube-controller-manager,--allocate-node-cidrs=true,--cloud-provider=,--cluster-cidr=,--service-cluster-ip-range=,--cluster-signing-cert-file=/system/secrets/kubernetes/kube-controller-manager/ca.crt,--cluster-signing-key-file=/system/secrets/kubernetes/kube-controller-manager/ca.key,--configure-cloud-routes=false,--kubeconfig=/system/secrets/kubernetes/kube-controller-manager/kubeconfig,--leader-elect=true,--root-ca-file=/system/secrets/kubernetes/kube-controller-manager/ca.crt,--service-account-private-key-file=/system/secrets/kubernetes/kube-controller-manager/service-account.key,--profiling=false 2021-03-09T13:59:34.293870359Z stderr F 2021/03/09 13:59:34 Now listening for interrupts 2021-03-09T13:59:34.761113762Z stdout F I0309 13:59:34.760982      10 serving.go:331] Generated self-signed cert in-memory

Checking controller runtime logs

Talos runs a set of controllers which work on resources to build and support Kubernetes control plane.

Some debugging information can be queried from the controller logs with talosctl logs controller-runtime:

$ talosctl -n <IP> logs controller-runtime 2021/03/09 13:57:11  secrets.EtcdController: controller starting 2021/03/09 13:57:11  config.MachineTypeController: controller starting 2021/03/09 13:57:11  k8s.ManifestApplyController: controller starting 2021/03/09 13:57:11  v1alpha1.BootstrapStatusController: controller starting 2021/03/09 13:57:11  v1alpha1.TimeStatusController: controller starting

Controllers run reconcile loop, so they might be starting, failing and restarting, that is expected behavior. Things to look for:

v1alpha1.BootstrapStatusController: bootkube initialized status not found: control plane is not self-hosted, running with static pods.

k8s.KubeletStaticPodController: writing static pod "/etc/kubernetes/manifests/talos-kube-apiserver.yaml": static pod definitions were rendered successfully.

k8s.ManifestApplyController: controller failed: error creating mapping for object /v1/Secret/bootstrap-token-q9pyzr: an error on the server ("") has prevented the request from succeeding: control plane endpoint is not up yet, bootstrap manifests can't be injected, controller is going to retry.

k8s.KubeletStaticPodController: controller failed: error refreshing pod status: error fetching pod status: an error on the server ("Authorization error (user=apiserver-kubelet-client, verb=get, resource=nodes, subresource=proxy)") has prevented the request from succeeding: kubelet hasn't been able to contact kube-apiserver yet to push pod status, controller is going to retry.

k8s.ManifestApplyController: created one of the bootstrap manifests got successfully applied.

secrets.KubernetesController: controller failed: missing cluster.aggregatorCA secret: Talos is running with 0.8 configuration, if the cluster was upgraded from 0.8, this is expected, and conversion process will fix machine config automatically. If this cluster was bootstrapped with version 0.9, machine configuration should be regenerated with 0.9 talosctl.

If there are no new messages in controller-runtime log, it means that controllers finished reconciling successfully.

Checking static pod definitions

Talos generates static pod definitions for kube-apiserver, kube-controller-manager, and kube-scheduler components based on machine configuration. These definitions can be checked as resources with talosctl get staticpods:

$ talosctl -n <IP> get staticpods -o yaml
get staticpods -o yaml
    namespace: controlplane
    id: kube-apiserver
    version: 2
    phase: running
        - k8s.StaticPodStatus("kube-apiserver")
    apiVersion: v1
    kind: Pod
        creationTimestamp: null
            k8s-app: kube-apiserver
            tier: control-plane
        name: kube-apiserver
        namespace: kube-system

Status of the static pods can queried with talosctl get staticpodstatus:

$ talosctl -n <IP> get staticpodstatus
NODE         NAMESPACE      TYPE              ID                                                           VERSION   READY   controlplane   StaticPodStatus   kube-system/kube-apiserver-talos-default-master-1            1         True   controlplane   StaticPodStatus   kube-system/kube-controller-manager-talos-default-master-1   1         True   controlplane   StaticPodStatus   kube-system/kube-scheduler-talos-default-master-1            1         True

Most important status is Ready printed as last column, complete status can be fetched by adding -o yaml flag.

Checking bootstrap manifests

As part of bootstrap process, Talos injects bootstrap manifests into Kubernetes API server. There are two kinds of manifests: system manifests built-in into Talos and extra manifests downloaded (custom CNI, extra manifests in the machine config):

$ talosctl -n <IP> get manifests --namespace=controlplane
NODE         NAMESPACE      TYPE       ID                               VERSION   controlplane   Manifest   00-kubelet-bootstrapping-token   1   controlplane   Manifest   01-csr-approver-role-binding     1   controlplane   Manifest   01-csr-node-bootstrap            1   controlplane   Manifest   01-csr-renewal-role-binding      1   controlplane   Manifest   02-kube-system-sa-role-binding   1   controlplane   Manifest   03-default-pod-security-policy   1   controlplane   Manifest   10-kube-proxy                    1   controlplane   Manifest   11-core-dns                      1   controlplane   Manifest   11-core-dns-svc                  1   controlplane   Manifest   11-kube-config-in-cluster        1
$ talosctl -n <IP> get manifests --namespace=extras
NODE         NAMESPACE   TYPE       ID                                                        VERSION   extras      Manifest   05-   1

Details of each manifests can be queried by adding -o yaml:

$ talosctl -n <IP> get manifests 01-csr-approver-role-binding --namespace=controlplane -o yaml
    namespace: controlplane
    id: 01-csr-approver-role-binding
    version: 1
    phase: running
    - apiVersion:
      kind: ClusterRoleBinding
        name: system-bootstrap-approve-node-client-csr
        kind: ClusterRole
        - apiGroup:
          kind: Group
          name: system:bootstrappers

Worker node is stuck with apid health check failures

Control plane nodes have enough secret material to generate apid server certificates, but worker nodes depend on control plane trustd services to generate certificates. Worker nodes wait for kubelet to join the cluster, then apid queries Kubernetes endpoints via control plane endpoint to find trustd endpoints, and use trustd to issue the certficiate.

So if apid health checks is failing on worker node:

  • make sure control plane endpoint is healthy
  • check that worker node kubelet joined the cluster