Troubleshooting Kubernetes

Posted on , Updated on , by Linode Community .

Traducciones al Español
Estamos traduciendo nuestros guías y tutoriales al Español. Es posible que usted esté viendo una traducción generada automáticamente. Estamos trabajando con traductores profesionales para verificar las traducciones de nuestro sitio web. Este proyecto es un trabajo en curso.
Try this guide to receive $100 at signup on a new account.
Marquee image for Troubleshooting Kubernetes

Troubleshooting issues with Kubernetes can be complex, and it can be difficult to account for all the possible error conditions you may see. This guide tries to equip you with the core tools that can be useful when troubleshooting, and it introduces some situations that you may find yourself in.

If your issue is not covered by this guide, we also recommend researching and posting in the Linode Community Questions site and in #linode on the Kubernetes Slack, where other Linode users (and the Kubernetes community) can offer advice.

If you are running a cluster on Linode’s managed LKE service, and you are experiencing an issue related to your master/control plane components, you can report these issues to Linode by contacting Linode Support. Examples in this category include:

  • Kubernetes' API server not running. If kubectl does not respond as expected, this can indicate problems with the API server.

  • The CCM, CSI, Calico, or kube-dns Pods are not running.

  • Annotations on LoadBalancer services aren’t functioning.

  • PersistentVolumes are not re-attaching.

Please note that the kube-apiserver and etcd Pods will not be visible for LKE clusters, and this is expected. Issues outside of the scope of Linode Support include:

  • Problems with the control plane of clusters not managed by LKE.

  • Problems with applications deployed on Kubernetes.

If your issue is not covered by this guide, we also recommend researching and posting in the Linode …

In this guide we will:

General Troubleshooting Strategies

To troubleshoot issues with the applications running on your cluster, you can rely on the kubectl command to gather debugging information. kubectl includes a set of subcommands that can be used to research issues with your cluster, and this guide will highlight four of them: get, describe, logs, and exec.

To troubleshoot issues with your cluster, you may need to directly view the logs that are generated by Kubernetes' components.

kubectl get

Use the get command to list different kinds of resources in your cluster (nodes, pods, services, etc). The output will show the status for each resource returned. For example, this output shows that a Pod is in the CrashLoopBackOff status, which means it should be investigated further:

kubectl get pods
NAME              READY     STATUS             RESTARTS   AGE
ghost-0           0/1       CrashLoopBackOff   34         2h
mariadb-0         1/1       Running            0          2h
  • Use the --namespace flag to show resources in a certain namespace:

    # Show pods in the `kube-system` namespace
    kubectl get pods --namespace kube-system
    If you’ve set up Kubernetes using automated solutions like Linode’s Kubernetes Engine, k8s-alpha CLI, or Rancher, you’ll see csi-linode and ccm-linode Pods in the kube-system namespace. This is normal as long as they’re in the Running status.
  • Use the -o flag to return the resources as YAML or JSON. The Kubernetes API’s complete description for the returned resources will be shown:

    # Get pods as YAML API objects
    kubectl get pods -o yaml
  • Sort the returned resources with the --sort-by flag:

    # Sort by name
    kubectl get pods
  • Use the --selector or -l flag to get resources that match a label. This is useful for finding all Pods for a given service:

    # Get pods which match the app=ghost selector
    kubectl get pods -l app=ghost
  • Use the --field-selector flag to return resources which match different resource fields:

    # Get all pods that are Pending
    kubectl get pods --field-selector status.phase=Pending
    # Get all pods that are not in the kube-system namespace
    kubectl get pods --field-selector metadata.namespace!=kube-system

kubectl describe

Use the describe command to return a detailed report of the state of one or more resources in your cluster. Pass a resource type to the describe command to get a report for each of those resources:

kubectl describe nodes

Pass the name of a resource to get a report for just that object:

kubectl describe pods ghost-0

You can also use the --selector (-l) flag to filter the returned resources, as with the get command.

kubectl logs

Use the logs command to print logs collected by a Pod:

kubectl logs mariadb-0
  • Use the --selector (-l) flag to print logs from all Pods that match a selector:

    kubectl logs -l app=ghost
  • If a Pod’s container was killed and restarted, you can view the previous container’s logs with the --previous or -p flag:

    kubectl logs -p ghost-0

kubectl exec

You can run arbitrary commands on a Pod’s container by passing them to kubectl’s exec command:

kubectl exec mariadb-0 -- ps aux

The full syntax for the command is:

kubectl exec ${POD_NAME} -c ${CONTAINER_NAME} -- ${CMD} ${ARG1} ${ARG2} ... ${ARGN}
The -c flag is optional, and is only needed when the specified Pod is running more than one container.

It is possible to run an interactive shell on an existing pod/container. Pass the -it flags to exec and run the shell:

kubectl exec -it mariadb-0 -- /bin/bash

Enter exit to later leave this shell.

Viewing Master and Worker Logs

If the Kubernetes API server isn’t working normally, then you may not be able to use kubectl to troubleshoot. When this happens, or if you are experiencing other more fundamental issues with your cluster, you can instead log directly into your nodes and view the logs present on your filesystem.

Non-systemd systems

If your nodes do not run systemd, the location of logs on your master nodes should be:

  • /var/log/kube-apiserver.log - API server

  • /var/log/kube-scheduler.log - Scheduler

  • /var/log/kube-controller-manager.log - Replication controller manager

    You will not be able to directly access the master nodes of your cluster if you are using Linode’s LKE service.

On your worker nodes:

systemd systems

If your nodes run systemd, you can access the logs that kubelet generates with journalctl:

journalctl --unit kubelet

Logs for your other Kubernetes software components can be found through your container runtime. When using Docker, you can use the docker ps and docker logs commands to investigate. For example, to find the container running your API server:

docker ps | grep apiserver

The output will display a list of information separated by tabs:

2f4e6242e1a2    cfdda15fbce2    "kube-apiserver --au…"  2 days ago  Up 2 days   k8s_kube-apiserver_kube-apiserver-k8s-trouble-1-master-1_kube-system_085b2ab3bd6d908bde1af92bd25e5aaa_0

The first entry (in this example: 2f4e6242e1a2) will be an alphanumeric string, and it is the ID for the container. Copy this string and pass it to docker logs to view the logs for your API server:

docker logs ${CONTAINER_ID}

Troubleshooting Examples

Viewing the Wrong Cluster

If your kubectl commands are not returning the resources and information you expect, then your client may be assigned to the wrong cluster context. To view all of the cluster contexts on your system, run:

kubectl config get-contexts

An asterisk will appear next to the active context:

CURRENT   NAME                                        CLUSTER            AUTHINFO
          [email protected]           my-cluster         my-cluster-kayciZBRO5s
*         [email protected]     other-cluster      other-cluster-kaUzJOMWJ3c

To switch to another context, run:

kubectl config use-context ${CLUSTER_NAME}


kubectl config use-context [email protected]

Can’t Provision Cluster Nodes

If you are not able to create new nodes in your cluster, you may see an error message similar to:

  Error creating a Linode Instance: [400] Account Limit reached. Please open a support ticket.

This is a reference to the total number of Linode resources that can exist on your account. To create new Linode instances for your cluster, you will need to either remove other instances on your account, or request a limit increase. To request a limit increase, contact Linode Support.

Insufficient CPU or Memory

If one of your Pods requests more memory or CPU than is available on your worker nodes, then one of these scenarios may happen:

  • The Pod will remain in the Pending state, because the scheduler cannot find a node to run it on. This will be visible when running kubectl get pods.

    If you run the kubectl describe command on your Pod, the Events section may list a FailedScheduling event, along with a message like Failed for reason PodExceedsFreeCPU and possibly others. You can run kubectl describe nodes to view information about the allocated resources for each node.

  • The Pod may continually crash. For example, the Ghost Pod specified by Ghost’s Helm chart will show the following error in its logs when not enough memory is available:

    kubectl logs ghost --tail=5
    1) SystemError
    Message: You are recommended to have at least 150 MB of memory available for smooth operation. It looks like you have ~58 MB available.

If your cluster has insufficient resources for a new Pod, you will need to:

  • Reduce the number of other pods/deployments/applications running on your cluster,
  • Resize the Linode instances that represent your worker nodes to a higher-tier plan, or
  • Add a new worker node to your cluster.

Your Feedback Is Important

Let us know if this guide helped you find the answer you were looking for.