はじめに

Edit This Page

kubeadmのトラブルシューティング

As with any program, you might run into an error installing or running kubeadm. This page lists some common failure scenarios and have provided steps that can help you understand and fix the problem.

If your problem is not listed below, please follow the following steps:

  • If you think your problem is a bug with kubeadm:

  • If you are unsure about how kubeadm works, you can ask on Slack in #kubeadm, or open a question on StackOverflow. Please include relevant tags like #kubernetes and #kubeadm so folks can help you.

インストール中にebtablesもしくは他の似たような実行プログラムが見つからない

If you see the following warnings while running kubeadm init

[preflight] WARNING: ebtables not found in system path
[preflight] WARNING: ethtool not found in system path

Then you may be missing ebtables, ethtool or a similar executable on your node. You can install them with the following commands:

  • For Ubuntu/Debian users, run apt install ebtables ethtool.
  • For CentOS/Fedora users, run yum install ebtables ethtool.

インストール中にkubeadmがコントロールプレーンを待ち続けて止まる

If you notice that kubeadm init hangs after printing out the following line:

[apiclient] Created API client, waiting for the control plane to become ready

This may be caused by a number of problems. The most common are:

  • network connection problems. Check that your machine has full network connectivity before continuing.
  • the default cgroup driver configuration for the kubelet differs from that used by Docker. Check the system log file (e.g. /var/log/message) or examine the output from journalctl -u kubelet. If you see something like the following:

    error: failed to run Kubelet: failed to create kubelet:
    misconfiguration: kubelet cgroup driver: "systemd" is different from docker cgroup driver: "cgroupfs"

There are two common ways to fix the cgroup driver problem:

  1. Install Docker again following instructions here.

  2. Change the kubelet config to match the Docker cgroup driver manually, you can refer to Configure cgroup driver used by kubelet on Master Node

  • control plane Docker containers are crashlooping or hanging. You can check this by running docker ps and investigating each container by running docker logs.

管理コンテナを削除する時にkubeadmが止まる

The following could happen if Docker halts and does not remove any Kubernetes-managed containers:

sudo kubeadm reset
[preflight] Running pre-flight checks
[reset] Stopping the kubelet service
[reset] Unmounting mounted directories in "/var/lib/kubelet"
[reset] Removing kubernetes-managed containers
(block)

A possible solution is to restart the Docker service and then re-run kubeadm reset:

sudo systemctl restart docker.service
sudo kubeadm reset

Inspecting the logs for docker may also be useful:

journalctl -ul docker

Podの状態がRunContainerErrorCrashLoopBackOff、またはError

Right after kubeadm init there should not be any pods in these states.

  • If there are pods in one of these states right after kubeadm init, please open an issue in the kubeadm repo. coredns (or kube-dns) should be in the Pending state until you have deployed the network solution.
  • If you see Pods in the RunContainerError, CrashLoopBackOff or Error state after deploying the network solution and nothing happens to coredns (or kube-dns), it’s very likely that the Pod Network solution that you installed is somehow broken. You might have to grant it more RBAC privileges or use a newer version. Please file an issue in the Pod Network providers’ issue tracker and get the issue triaged there.
  • If you install a version of Docker older than 1.12.1, remove the MountFlags=slave option when booting dockerd with systemd and restart docker. You can see the MountFlags in /usr/lib/systemd/system/docker.service. MountFlags can interfere with volumes mounted by Kubernetes, and put the Pods in CrashLoopBackOff state. The error happens when Kubernetes does not find var/run/secrets/kubernetes.io/serviceaccount files.

coredns(もしくはkube-dns)がPending状態でスタックする

This is expected and part of the design. kubeadm is network provider-agnostic, so the admin should install the pod network solution of choice. You have to install a Pod Network before CoreDNS may be deployed fully. Hence the Pending state before the network is set up.

HostPortサービスが動かない

The HostPort and HostIP functionality is available depending on your Pod Network provider. Please contact the author of the Pod Network solution to find out whether HostPort and HostIP functionality are available.

Calico, Canal, and Flannel CNI providers are verified to support HostPort.

For more information, see the CNI portmap documentation.

If your network provider does not support the portmap CNI plugin, you may need to use the NodePort feature of services or use HostNetwork=true.

サービスIP経由でPodにアクセスすることができない

  • Many network add-ons do not yet enable hairpin mode which allows pods to access themselves via their Service IP. This is an issue related to CNI. Please contact the network add-on provider to get the latest status of their support for hairpin mode.

  • If you are using VirtualBox (directly or via Vagrant), you will need to ensure that hostname -i returns a routable IP address. By default the first interface is connected to a non-routable host-only network. A work around is to modify /etc/hosts, see this Vagrantfile for an example.

TLS証明書のエラー

The following error indicates a possible certificate mismatch.

# kubectl get pods
Unable to connect to the server: x509: certificate signed by unknown authority (possibly because of "crypto/rsa: verification error" while trying to verify candidate authority certificate "kubernetes")
  • Verify that the $HOME/.kube/config file contains a valid certificate, and regenerate a certificate if necessary. The certificates in a kubeconfig file are base64 encoded. The base64 -d command can be used to decode the certificate and openssl x509 -text -noout can be used for viewing the certificate information.
  • Unset the KUBECONFIG environment variable using:

    unset KUBECONFIG

Or set it to the default KUBECONFIG location:

  export KUBECONFIG=/etc/kubernetes/admin.conf
  • Another workaround is to overwrite the existing kubeconfig for the “admin” user:

    mv  $HOME/.kube $HOME/.kube.bak
    sudo cp -i /etc/kubernetes/admin.conf $HOME/.kube/config
    sudo chown $(id -u):$(id -g) $HOME/.kube/config

Vagrant内でPodネットワークとしてflannelを使用する時のデフォルトNIC

The following error might indicate that something was wrong in the pod network:

Error from server (NotFound): the server could not find the requested resource
  • If you’re using flannel as the pod network inside Vagrant, then you will have to specify the default interface name for flannel.

Vagrant typically assigns two interfaces to all VMs. The first, for which all hosts are assigned the IP address 10.0.2.15, is for external traffic that gets NATed.

This may lead to problems with flannel, which defaults to the first interface on a host. This leads to all hosts thinking they have the same public IP address. To prevent this, pass the --iface eth1 flag to flannel so that the second interface is chosen.

公開されていないIPがコンテナに使われている

In some situations kubectl logs and kubectl run commands may return with the following errors in an otherwise functional cluster:

Error from server: Get https://10.19.0.41:10250/containerLogs/default/mysql-ddc65b868-glc5m/mysql: dial tcp 10.19.0.41:10250: getsockopt: no route to host
  • This may be due to Kubernetes using an IP that can not communicate with other IPs on the seemingly same subnet, possibly by policy of the machine provider.
  • Digital Ocean assigns a public IP to eth0 as well as a private one to be used internally as anchor for their floating IP feature, yet kubelet will pick the latter as the node’s InternalIP instead of the public one.

Use ip addr show to check for this scenario instead of ifconfig because ifconfig will not display the offending alias IP address. Alternatively an API endpoint specific to Digital Ocean allows to query for the anchor IP from the droplet:

  curl http://169.254.169.254/metadata/v1/interfaces/public/0/anchor_ipv4/address

The workaround is to tell kubelet which IP to use using --node-ip. When using Digital Ocean, it can be the public one (assigned to eth0) or the private one (assigned to eth1) should you want to use the optional private network. The KubeletExtraArgs section of the kubeadm NodeRegistrationOptions structure can be used for this.

Then restart kubelet:

  systemctl daemon-reload
  systemctl restart kubelet

corednsのPodがCrashLoopBackOffもしくはError状態になる

If you have nodes that are running SELinux with an older version of Docker you might experience a scenario where the coredns pods are not starting. To solve that you can try one of the following options:

  • Upgrade to a newer version of Docker.

  • Disable SELinux.

  • Modify the coredns deployment to set allowPrivilegeEscalation to true:

    kubectl -n kube-system get deployment coredns -o yaml | \
    sed 's/allowPrivilegeEscalation: false/allowPrivilegeEscalation: true/g' | \
    kubectl apply -f -

Another cause for CoreDNS to have CrashLoopBackOff is when a CoreDNS Pod deployed in Kubernetes detects a loop. A number of workarounds are available to avoid Kubernetes trying to restart the CoreDNS Pod every time CoreDNS detects the loop and exits.

警告: Disabling SELinux or setting allowPrivilegeEscalation to true can compromise the security of your cluster.

etcdのpodが継続的に再起動する

If you encounter the following error:

rpc error: code = 2 desc = oci runtime error: exec failed: container_linux.go:247: starting container process caused "process_linux.go:110: decoding init error from pipe caused \"read parent: connection reset by peer\""

this issue appears if you run CentOS 7 with Docker 1.13.1.84. This version of Docker can prevent the kubelet from executing into the etcd container.

To work around the issue, choose one of these options:

  • Roll back to an earlier version of Docker, such as 1.13.1-75

    yum downgrade docker-1.13.1-75.git8633870.el7.centos.x86_64 docker-client-1.13.1-75.git8633870.el7.centos.x86_64 docker-common-1.13.1-75.git8633870.el7.centos.x86_64
    
  • Install one of the more recent recommended versions, such as 18.06:

    sudo yum-config-manager --add-repo https://download.docker.com/linux/centos/docker-ce.repo
    yum install docker-ce-18.06.1.ce-3.el7.x86_64

Not possible to pass a comma separated list of values to arguments inside a --component-extra-args flag

kubeadm init flags such as --component-extra-args allow you to pass custom arguments to a control-plane component like the kube-apiserver. However, this mechanism is limited due to the underlying type used for parsing the values (mapStringString).

If you decide to pass an argument that supports multiple, comma-separated values such as --apiserver-extra-args "enable-admission-plugins=LimitRanger,NamespaceExists" this flag will fail with flag: malformed pair, expect string=string. This happens because the list of arguments for --apiserver-extra-args expects key=value pairs and in this case NamespacesExists is considered as a key that is missing a value.

Alternatively, you can try separating the key=value pairs like so: --apiserver-extra-args "enable-admission-plugins=LimitRanger,enable-admission-plugins=NamespaceExists" but this will result in the key enable-admission-plugins only having the value of NamespaceExists.

A known workaround is to use the kubeadm configuration file.

kube-proxy scheduled before node is initialized by cloud-controller-manager

In cloud provider scenarios, kube-proxy can end up being scheduled on new worker nodes before the cloud-controller-manager has initialized the node addresses. This causes kube-proxy to fail to pick up the node’s IP address properly and has knock-on effects to the proxy function managing load balancers.

The following error can be seen in kube-proxy Pods:

server.go:610] Failed to retrieve node IP: host IP unknown; known addresses: []
proxier.go:340] invalid nodeIP, initializing kube-proxy with 127.0.0.1 as nodeIP

A known solution is to patch the kube-proxy DaemonSet to allow scheduling it on control-plane nodes regardless of their conditions, keeping it off of other nodes until their initial guarding conditions abate:

kubectl -n kube-system patch ds kube-proxy -p='{ "spec": { "template": { "spec": { "tolerations": [ { "key": "CriticalAddonsOnly", "operator": "Exists" }, { "effect": "NoSchedule", "key": "node-role.kubernetes.io/master" } ] } } } }'

The tracking issue for this problem is here.

The NodeRegistration.Taints field is omitted when marshalling kubeadm configuration

Note: This issue only applies to tools that marshal kubeadm types (e.g. to a YAML configuration file). It will be fixed in kubeadm API v1beta2.

By default, kubeadm applies the role.kubernetes.io/master:NoSchedule taint to control-plane nodes. If you prefer kubeadm to not taint the control-plane node, and set InitConfiguration.NodeRegistration.Taints to an empty slice, the field will be omitted when marshalling. When the field is omitted, kubeadm applies the default taint.

There are at least two workarounds:

  1. Use the role.kubernetes.io/master:PreferNoSchedule taint instead of an empty slice. Pods will get scheduled on masters, unless other nodes have capacity.

  2. Remove the taint after kubeadm init exits:

    kubectl taint nodes NODE_NAME role.kubernetes.io/master:NoSchedule-

フィードバック