1 - Container runtimes

You need to install a container runtime into each node in the cluster so that Pods can run there. This page outlines what is involved and describes related tasks for setting up nodes.

This page lists details for using several common container runtimes with Kubernetes, on Linux:

Note: For other operating systems, look for documentation specific to your platform.

Cgroup drivers

Control groups are used to constrain resources that are allocated to processes.

When systemd is chosen as the init system for a Linux distribution, the init process generates and consumes a root control group (cgroup) and acts as a cgroup manager. Systemd has a tight integration with cgroups and allocates a cgroup per systemd unit. It's possible to configure your container runtime and the kubelet to use cgroupfs. Using cgroupfs alongside systemd means that there will be two different cgroup managers.

A single cgroup manager simplifies the view of what resources are being allocated and will by default have a more consistent view of the available and in-use resources. When there are two cgroup managers on a system, you end up with two views of those resources. In the field, people have reported cases where nodes that are configured to use cgroupfs for the kubelet and Docker, but systemd for the rest of the processes, become unstable under resource pressure.

Changing the settings such that your container runtime and kubelet use systemd as the cgroup driver stabilized the system. To configure this for Docker, set native.cgroupdriver=systemd.


Changing the cgroup driver of a Node that has joined a cluster is a sensitive operation. If the kubelet has created Pods using the semantics of one cgroup driver, changing the container runtime to another cgroup driver can cause errors when trying to re-create the Pod sandbox for such existing Pods. Restarting the kubelet may not solve such errors.

If you have automation that makes it feasible, replace the node with another using the updated configuration, or reinstall it using automation.

Migrating to the systemd driver in kubeadm managed clusters

Follow this Migration guide if you wish to migrate to the systemd cgroup driver in existing kubeadm managed clusters.

Container runtimes


This section contains the necessary steps to use containerd as CRI runtime.

Use the following commands to install Containerd on your system:

Install and configure prerequisites:

cat <<EOF | sudo tee /etc/modules-load.d/containerd.conf

sudo modprobe overlay
sudo modprobe br_netfilter

# Setup required sysctl params, these persist across reboots.
cat <<EOF | sudo tee /etc/sysctl.d/99-kubernetes-cri.conf
net.bridge.bridge-nf-call-iptables  = 1
net.ipv4.ip_forward                 = 1
net.bridge.bridge-nf-call-ip6tables = 1

# Apply sysctl params without reboot
sudo sysctl --system

Install containerd:

  1. Install the containerd.io package from the official Docker repositories. Instructions for setting up the Docker repository for your respective Linux distribution and installing the containerd.io package can be found at Install Docker Engine.

  2. Configure containerd:

    sudo mkdir -p /etc/containerd
    containerd config default | sudo tee /etc/containerd/config.toml
  3. Restart containerd:

    sudo systemctl restart containerd

Start a Powershell session, set $Version to the desired version (ex: $Version=1.4.3), and then run the following commands:

  1. Download containerd:

    curl.exe -L https://github.com/containerd/containerd/releases/download/v$Version/containerd-$Version-windows-amd64.tar.gz -o containerd-windows-amd64.tar.gz
    tar.exe xvf .\containerd-windows-amd64.tar.gz
  2. Extract and configure:

    Copy-Item -Path ".\bin\" -Destination "$Env:ProgramFiles\containerd" -Recurse -Force
    cd $Env:ProgramFiles\containerd\
    .\containerd.exe config default | Out-File config.toml -Encoding ascii
    # Review the configuration. Depending on setup you may want to adjust:
    # - the sandbox_image (Kubernetes pause image)
    # - cni bin_dir and conf_dir locations
    Get-Content config.toml
    # (Optional - but highly recommended) Exclude containerd from Windows Defender Scans
    Add-MpPreference -ExclusionProcess "$Env:ProgramFiles\containerd\containerd.exe"
  3. Start containerd:

    .\containerd.exe --register-service
    Start-Service containerd

Using the systemd cgroup driver

To use the systemd cgroup driver in /etc/containerd/config.toml with runc, set

    SystemdCgroup = true

If you apply this change make sure to restart containerd again:

sudo systemctl restart containerd

When using kubeadm, manually configure the cgroup driver for kubelet.


This section contains the necessary steps to install CRI-O as a container runtime.

Use the following commands to install CRI-O on your system:

Note: The CRI-O major and minor versions must match the Kubernetes major and minor versions. For more information, see the CRI-O compatibility matrix.

Install and configure prerequisites:

# Create the .conf file to load the modules at bootup
cat <<EOF | sudo tee /etc/modules-load.d/crio.conf

sudo modprobe overlay
sudo modprobe br_netfilter

# Set up required sysctl params, these persist across reboots.
cat <<EOF | sudo tee /etc/sysctl.d/99-kubernetes-cri.conf
net.bridge.bridge-nf-call-iptables  = 1
net.ipv4.ip_forward                 = 1
net.bridge.bridge-nf-call-ip6tables = 1

sudo sysctl --system

To install CRI-O on the following operating systems, set the environment variable OS to the appropriate value from the following table:

Operating system$OS
Debian UnstableDebian_Unstable
Debian TestingDebian_Testing

Then, set $VERSION to the CRI-O version that matches your Kubernetes version. For instance, if you want to install CRI-O 1.20, set VERSION=1.20. You can pin your installation to a specific release. To install version 1.20.0, set VERSION=1.20:1.20.0.

Then run

cat <<EOF | sudo tee /etc/apt/sources.list.d/devel:kubic:libcontainers:stable.list
deb https://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable/$OS/ /
cat <<EOF | sudo tee /etc/apt/sources.list.d/devel:kubic:libcontainers:stable:cri-o:$VERSION.list
deb http://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable:/cri-o:/$VERSION/$OS/ /

curl -L https://download.opensuse.org/repositories/devel:kubic:libcontainers:stable:cri-o:$VERSION/$OS/Release.key | sudo apt-key --keyring /etc/apt/trusted.gpg.d/libcontainers.gpg add -
curl -L https://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable/$OS/Release.key | sudo apt-key --keyring /etc/apt/trusted.gpg.d/libcontainers.gpg add -

sudo apt-get update
sudo apt-get install cri-o cri-o-runc

To install on the following operating systems, set the environment variable OS to the appropriate field in the following table:

Operating system$OS
Ubuntu 20.04xUbuntu_20.04
Ubuntu 19.10xUbuntu_19.10
Ubuntu 19.04xUbuntu_19.04
Ubuntu 18.04xUbuntu_18.04

Then, set $VERSION to the CRI-O version that matches your Kubernetes version. For instance, if you want to install CRI-O 1.20, set VERSION=1.20. You can pin your installation to a specific release. To install version 1.20.0, set VERSION=1.20:1.20.0.

Then run

cat <<EOF | sudo tee /etc/apt/sources.list.d/devel:kubic:libcontainers:stable.list
deb https://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable/$OS/ /
cat <<EOF | sudo tee /etc/apt/sources.list.d/devel:kubic:libcontainers:stable:cri-o:$VERSION.list
deb http://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable:/cri-o:/$VERSION/$OS/ /

curl -L https://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable/$OS/Release.key | sudo apt-key --keyring /etc/apt/trusted.gpg.d/libcontainers.gpg add -
curl -L https://download.opensuse.org/repositories/devel:kubic:libcontainers:stable:cri-o:$VERSION/$OS/Release.key | sudo apt-key --keyring /etc/apt/trusted.gpg.d/libcontainers-cri-o.gpg add -

sudo apt-get update
sudo apt-get install cri-o cri-o-runc

To install on the following operating systems, set the environment variable OS to the appropriate field in the following table:

Operating system$OS
Centos 8CentOS_8
Centos 8 StreamCentOS_8_Stream
Centos 7CentOS_7

Then, set $VERSION to the CRI-O version that matches your Kubernetes version. For instance, if you want to install CRI-O 1.20, set VERSION=1.20. You can pin your installation to a specific release. To install version 1.20.0, set VERSION=1.20:1.20.0.

Then run

sudo curl -L -o /etc/yum.repos.d/devel:kubic:libcontainers:stable.repo https://download.opensuse.org/repositories/devel:/kubic:/libcontainers:/stable/$OS/devel:kubic:libcontainers:stable.repo
sudo curl -L -o /etc/yum.repos.d/devel:kubic:libcontainers:stable:cri-o:$VERSION.repo https://download.opensuse.org/repositories/devel:kubic:libcontainers:stable:cri-o:$VERSION/$OS/devel:kubic:libcontainers:stable:cri-o:$VERSION.repo
sudo yum install cri-o

sudo zypper install cri-o

Set $VERSION to the CRI-O version that matches your Kubernetes version. For instance, if you want to install CRI-O 1.20, VERSION=1.20.

You can find available versions with:

sudo dnf module list cri-o

CRI-O does not support pinning to specific releases on Fedora.

Then run

sudo dnf module enable cri-o:$VERSION
sudo dnf install cri-o

Start CRI-O:

sudo systemctl daemon-reload
sudo systemctl enable crio --now

Refer to the CRI-O installation guide for more information.

cgroup driver

CRI-O uses the systemd cgroup driver per default. To switch to the cgroupfs cgroup driver, either edit /etc/crio/crio.conf or place a drop-in configuration in /etc/crio/crio.conf.d/02-cgroup-manager.conf, for example:

conmon_cgroup = "pod"
cgroup_manager = "cgroupfs"

Please also note the changed conmon_cgroup, which has to be set to the value pod when using CRI-O with cgroupfs. It is generally necessary to keep the cgroup driver configuration of the kubelet (usually done via kubeadm) and CRI-O in sync.


  1. On each of your nodes, install the Docker for your Linux distribution as per Install Docker Engine. You can find the latest validated version of Docker in this dependencies file.

  2. Configure the Docker daemon, in particular to use systemd for the management of the container’s cgroups.

    sudo mkdir /etc/docker
    cat <<EOF | sudo tee /etc/docker/daemon.json
      "exec-opts": ["native.cgroupdriver=systemd"],
      "log-driver": "json-file",
      "log-opts": {
        "max-size": "100m"
      "storage-driver": "overlay2"
    Note: overlay2 is the preferred storage driver for systems running Linux kernel version 4.0 or higher, or RHEL or CentOS using version 3.10.0-514 and above.
  3. Restart Docker and enable on boot:

    sudo systemctl enable docker
    sudo systemctl daemon-reload
    sudo systemctl restart docker

2 - Installing Kubernetes with deployment tools

2.1 - Bootstrapping clusters with kubeadm

2.1.1 - Installing kubeadm

This page shows how to install the kubeadm toolbox. For information how to create a cluster with kubeadm once you have performed this installation process, see the Using kubeadm to Create a Cluster page.

Before you begin

  • A compatible Linux host. The Kubernetes project provides generic instructions for Linux distributions based on Debian and Red Hat, and those distributions without a package manager.
  • 2 GB or more of RAM per machine (any less will leave little room for your apps).
  • 2 CPUs or more.
  • Full network connectivity between all machines in the cluster (public or private network is fine).
  • Unique hostname, MAC address, and product_uuid for every node. See here for more details.
  • Certain ports are open on your machines. See here for more details.
  • Swap disabled. You MUST disable swap in order for the kubelet to work properly.

Verify the MAC address and product_uuid are unique for every node

  • You can get the MAC address of the network interfaces using the command ip link or ifconfig -a
  • The product_uuid can be checked by using the command sudo cat /sys/class/dmi/id/product_uuid

It is very likely that hardware devices will have unique addresses, although some virtual machines may have identical values. Kubernetes uses these values to uniquely identify the nodes in the cluster. If these values are not unique to each node, the installation process may fail.

Check network adapters

If you have more than one network adapter, and your Kubernetes components are not reachable on the default route, we recommend you add IP route(s) so Kubernetes cluster addresses go via the appropriate adapter.

Letting iptables see bridged traffic

Make sure that the br_netfilter module is loaded. This can be done by running lsmod | grep br_netfilter. To load it explicitly call sudo modprobe br_netfilter.

As a requirement for your Linux Node's iptables to correctly see bridged traffic, you should ensure net.bridge.bridge-nf-call-iptables is set to 1 in your sysctl config, e.g.

cat <<EOF | sudo tee /etc/modules-load.d/k8s.conf

cat <<EOF | sudo tee /etc/sysctl.d/k8s.conf
net.bridge.bridge-nf-call-ip6tables = 1
net.bridge.bridge-nf-call-iptables = 1
sudo sysctl --system

For more details please see the Network Plugin Requirements page.

Check required ports

Control-plane node(s)

ProtocolDirectionPort RangePurposeUsed By
TCPInbound6443*Kubernetes API serverAll
TCPInbound2379-2380etcd server client APIkube-apiserver, etcd
TCPInbound10250kubelet APISelf, Control plane

Worker node(s)

ProtocolDirectionPort RangePurposeUsed By
TCPInbound10250kubelet APISelf, Control plane
TCPInbound30000-32767NodePort Services†All

† Default port range for NodePort Services.

Any port numbers marked with * are overridable, so you will need to ensure any custom ports you provide are also open.

Although etcd ports are included in control-plane nodes, you can also host your own etcd cluster externally or on custom ports.

The pod network plugin you use (see below) may also require certain ports to be open. Since this differs with each pod network plugin, please see the documentation for the plugins about what port(s) those need.

Installing runtime

To run containers in Pods, Kubernetes uses a container runtime.

By default, Kubernetes uses the Container Runtime Interface (CRI) to interface with your chosen container runtime.

If you don't specify a runtime, kubeadm automatically tries to detect an installed container runtime by scanning through a list of well known Unix domain sockets. The following table lists container runtimes and their associated socket paths:

Container runtimes and their socket paths
RuntimePath to Unix domain socket

If both Docker and containerd are detected, Docker takes precedence. This is needed because Docker 18.09 ships with containerd and both are detectable even if you only installed Docker. If any other two or more runtimes are detected, kubeadm exits with an error.

The kubelet integrates with Docker through the built-in dockershim CRI implementation.

See container runtimes for more information.

By default, kubeadm uses Docker as the container runtime. The kubelet integrates with Docker through the built-in dockershim CRI implementation.

See container runtimes for more information.

Installing kubeadm, kubelet and kubectl

You will install these packages on all of your machines:

  • kubeadm: the command to bootstrap the cluster.

  • kubelet: the component that runs on all of the machines in your cluster and does things like starting pods and containers.

  • kubectl: the command line util to talk to your cluster.

kubeadm will not install or manage kubelet or kubectl for you, so you will need to ensure they match the version of the Kubernetes control plane you want kubeadm to install for you. If you do not, there is a risk of a version skew occurring that can lead to unexpected, buggy behaviour. However, one minor version skew between the kubelet and the control plane is supported, but the kubelet version may never exceed the API server version. For example, the kubelet running 1.7.0 should be fully compatible with a 1.8.0 API server, but not vice versa.

For information about installing kubectl, see Install and set up kubectl.

Warning: These instructions exclude all Kubernetes packages from any system upgrades. This is because kubeadm and Kubernetes require special attention to upgrade.

For more information on version skews, see:

  1. Update the apt package index and install packages needed to use the Kubernetes apt repository:

    sudo apt-get update
    sudo apt-get install -y apt-transport-https ca-certificates curl
  2. Download the Google Cloud public signing key:

    sudo curl -fsSLo /usr/share/keyrings/kubernetes-archive-keyring.gpg https://packages.cloud.google.com/apt/doc/apt-key.gpg
  3. Add the Kubernetes apt repository:

    echo "deb [signed-by=/usr/share/keyrings/kubernetes-archive-keyring.gpg] https://apt.kubernetes.io/ kubernetes-xenial main" | sudo tee /etc/apt/sources.list.d/kubernetes.list
  4. Update apt package index, install kubelet, kubeadm and kubectl, and pin their version:

    sudo apt-get update
    sudo apt-get install -y kubelet kubeadm kubectl
    sudo apt-mark hold kubelet kubeadm kubectl

cat <<EOF | sudo tee /etc/yum.repos.d/kubernetes.repo
gpgkey=https://packages.cloud.google.com/yum/doc/yum-key.gpg https://packages.cloud.google.com/yum/doc/rpm-package-key.gpg
exclude=kubelet kubeadm kubectl

# Set SELinux in permissive mode (effectively disabling it)
sudo setenforce 0
sudo sed -i 's/^SELINUX=enforcing$/SELINUX=permissive/' /etc/selinux/config

sudo yum install -y kubelet kubeadm kubectl --disableexcludes=kubernetes

sudo systemctl enable --now kubelet


  • Setting SELinux in permissive mode by running setenforce 0 and sed ... effectively disables it. This is required to allow containers to access the host filesystem, which is needed by pod networks for example. You have to do this until SELinux support is improved in the kubelet.

  • You can leave SELinux enabled if you know how to configure it but it may require settings that are not supported by kubeadm.

Install CNI plugins (required for most pod network):

sudo mkdir -p /opt/cni/bin
curl -L "https://github.com/containernetworking/plugins/releases/download/${CNI_VERSION}/cni-plugins-linux-amd64-${CNI_VERSION}.tgz" | sudo tar -C /opt/cni/bin -xz

Define the directory to download command files

Note: The DOWNLOAD_DIR variable must be set to a writable directory. If you are running Flatcar Container Linux, set DOWNLOAD_DIR=/opt/bin.
sudo mkdir -p $DOWNLOAD_DIR

Install crictl (required for kubeadm / Kubelet Container Runtime Interface (CRI))

curl -L "https://github.com/kubernetes-sigs/cri-tools/releases/download/${CRICTL_VERSION}/crictl-${CRICTL_VERSION}-linux-amd64.tar.gz" | sudo tar -C $DOWNLOAD_DIR -xz

Install kubeadm, kubelet, kubectl and add a kubelet systemd service:

RELEASE="$(curl -sSL https://dl.k8s.io/release/stable.txt)"
sudo curl -L --remote-name-all https://storage.googleapis.com/kubernetes-release/release/${RELEASE}/bin/linux/amd64/{kubeadm,kubelet,kubectl}
sudo chmod +x {kubeadm,kubelet,kubectl}

curl -sSL "https://raw.githubusercontent.com/kubernetes/release/${RELEASE_VERSION}/cmd/kubepkg/templates/latest/deb/kubelet/lib/systemd/system/kubelet.service" | sed "s:/usr/bin:${DOWNLOAD_DIR}:g" | sudo tee /etc/systemd/system/kubelet.service
sudo mkdir -p /etc/systemd/system/kubelet.service.d
curl -sSL "https://raw.githubusercontent.com/kubernetes/release/${RELEASE_VERSION}/cmd/kubepkg/templates/latest/deb/kubeadm/10-kubeadm.conf" | sed "s:/usr/bin:${DOWNLOAD_DIR}:g" | sudo tee /etc/systemd/system/kubelet.service.d/10-kubeadm.conf

Enable and start kubelet:

systemctl enable --now kubelet
Note: The Flatcar Container Linux distribution mounts the /usr directory as a read-only filesystem. Before bootstrapping your cluster, you need to take additional steps to configure a writable directory. See the Kubeadm Troubleshooting guide to learn how to set up a writable directory.

The kubelet is now restarting every few seconds, as it waits in a crashloop for kubeadm to tell it what to do.

Configuring a cgroup driver

Both the container runtime and the kubelet have a property called "cgroup driver", which is important for the management of cgroups on Linux machines.


Matching the container runtime and kubelet cgroup drivers is required or otherwise the kubelet process will fail.

See Configuring a cgroup driver for more details.


If you are running into difficulties with kubeadm, please consult our troubleshooting docs.

What's next

2.1.2 - Troubleshooting kubeadm

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.

Not possible to join a v1.18 Node to a v1.17 cluster due to missing RBAC

In v1.18 kubeadm added prevention for joining a Node in the cluster if a Node with the same name already exists. This required adding RBAC for the bootstrap-token user to be able to GET a Node object.

However this causes an issue where kubeadm join from v1.18 cannot join a cluster created by kubeadm v1.17.

To workaround the issue you have two options:

Execute kubeadm init phase bootstrap-token on a control-plane node using kubeadm v1.18. Note that this enables the rest of the bootstrap-token permissions as well.


Apply the following RBAC manually using kubectl apply -f ...:

apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
  name: kubeadm:get-nodes
- apiGroups:
  - ""
  - nodes
  - get
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
  name: kubeadm:get-nodes
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: kubeadm:get-nodes
- apiGroup: rbac.authorization.k8s.io
  kind: Group
  name: system:bootstrappers:kubeadm:default-node-token

ebtables or some similar executable not found during installation

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.

kubeadm blocks waiting for control plane during installation

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 control-plane 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.

kubeadm blocks when removing managed containers

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

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 -u docker

Pods in RunContainerError, CrashLoopBackOff or Error state

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 add-on.
  • If you see Pods in the RunContainerError, CrashLoopBackOff or Error state after deploying the network add-on and nothing happens to coredns (or kube-dns), it's very likely that the Pod Network add-on 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 is stuck in the Pending state

This is expected and part of the design. kubeadm is network provider-agnostic, so the admin should install the pod network add-on 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 services do not work

The HostPort and HostIP functionality is available depending on your Pod Network provider. Please contact the author of the Pod Network add-on 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.

Pods are not accessible via their Service IP

  • 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 certificate errors

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

Default NIC When using flannel as the pod network in Vagrant

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, 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.

Non-public IP used for containers

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

Error from server: Get dial tcp 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.

  • DigitalOcean 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 DigitalOcean allows to query for the anchor IP from the droplet:


    The workaround is to tell kubelet which IP to use using --node-ip. When using DigitalOcean, 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 pods have CrashLoopBackOff or Error state

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:

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.

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

etcd pods restart continually

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 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 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 node-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 node-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 node-role.kubernetes.io/master:NoSchedule-

/usr is mounted read-only on nodes

On Linux distributions such as Fedora CoreOS or Flatcar Container Linux, the directory /usr is mounted as a read-only filesystem. For flex-volume support, Kubernetes components like the kubelet and kube-controller-manager use the default path of /usr/libexec/kubernetes/kubelet-plugins/volume/exec/, yet the flex-volume directory must be writeable for the feature to work.

To workaround this issue you can configure the flex-volume directory using the kubeadm configuration file.

On the primary control-plane Node (created using kubeadm init) pass the following file using --config:

apiVersion: kubeadm.k8s.io/v1beta2
kind: InitConfiguration
    volume-plugin-dir: "/opt/libexec/kubernetes/kubelet-plugins/volume/exec/"
apiVersion: kubeadm.k8s.io/v1beta2
kind: ClusterConfiguration
    flex-volume-plugin-dir: "/opt/libexec/kubernetes/kubelet-plugins/volume/exec/"

On joining Nodes:

apiVersion: kubeadm.k8s.io/v1beta2
kind: JoinConfiguration
    volume-plugin-dir: "/opt/libexec/kubernetes/kubelet-plugins/volume/exec/"

Alternatively, you can modify /etc/fstab to make the /usr mount writeable, but please be advised that this is modifying a design principle of the Linux distribution.

kubeadm upgrade plan prints out context deadline exceeded error message

This error message is shown when upgrading a Kubernetes cluster with kubeadm in the case of running an external etcd. This is not a critical bug and happens because older versions of kubeadm perform a version check on the external etcd cluster. You can proceed with kubeadm upgrade apply ....

This issue is fixed as of version 1.19.

kubeadm reset unmounts /var/lib/kubelet

If /var/lib/kubelet is being mounted, performing a kubeadm reset will effectively unmount it.

To workaround the issue, re-mount the /var/lib/kubelet directory after performing the kubeadm reset operation.

This is a regression introduced in kubeadm 1.15. The issue is fixed in 1.20.

2.1.3 - Creating a cluster with kubeadm

Using kubeadm, you can create a minimum viable Kubernetes cluster that conforms to best practices. In fact, you can use kubeadm to set up a cluster that will pass the Kubernetes Conformance tests. kubeadm also supports other cluster lifecycle functions, such as bootstrap tokens and cluster upgrades.

The kubeadm tool is good if you need:

  • A simple way for you to try out Kubernetes, possibly for the first time.
  • A way for existing users to automate setting up a cluster and test their application.
  • A building block in other ecosystem and/or installer tools with a larger scope.

You can install and use kubeadm on various machines: your laptop, a set of cloud servers, a Raspberry Pi, and more. Whether you're deploying into the cloud or on-premises, you can integrate kubeadm into provisioning systems such as Ansible or Terraform.

Before you begin

To follow this guide, you need:

  • One or more machines running a deb/rpm-compatible Linux OS; for example: Ubuntu or CentOS.
  • 2 GiB or more of RAM per machine--any less leaves little room for your apps.
  • At least 2 CPUs on the machine that you use as a control-plane node.
  • Full network connectivity among all machines in the cluster. You can use either a public or a private network.

You also need to use a version of kubeadm that can deploy the version of Kubernetes that you want to use in your new cluster.

Kubernetes' version and version skew support policy applies to kubeadm as well as to Kubernetes overall. Check that policy to learn about what versions of Kubernetes and kubeadm are supported. This page is written for Kubernetes v1.21.

The kubeadm tool's overall feature state is General Availability (GA). Some sub-features are still under active development. The implementation of creating the cluster may change slightly as the tool evolves, but the overall implementation should be pretty stable.

Note: Any commands under kubeadm alpha are, by definition, supported on an alpha level.


  • Install a single control-plane Kubernetes cluster
  • Install a Pod network on the cluster so that your Pods can talk to each other


Installing kubeadm on your hosts

See "Installing kubeadm".


If you have already installed kubeadm, run apt-get update && apt-get upgrade or yum update to get the latest version of kubeadm.

When you upgrade, the kubelet restarts every few seconds as it waits in a crashloop for kubeadm to tell it what to do. This crashloop is expected and normal. After you initialize your control-plane, the kubelet runs normally.

Initializing your control-plane node

The control-plane node is the machine where the control plane components run, including etcd (the cluster database) and the API Server (which the kubectl command line tool communicates with).

  1. (Recommended) If you have plans to upgrade this single control-plane kubeadm cluster to high availability you should specify the --control-plane-endpoint to set the shared endpoint for all control-plane nodes. Such an endpoint can be either a DNS name or an IP address of a load-balancer.
  2. Choose a Pod network add-on, and verify whether it requires any arguments to be passed to kubeadm init. Depending on which third-party provider you choose, you might need to set the --pod-network-cidr to a provider-specific value. See Installing a Pod network add-on.
  3. (Optional) Since version 1.14, kubeadm tries to detect the container runtime on Linux by using a list of well known domain socket paths. To use different container runtime or if there are more than one installed on the provisioned node, specify the --cri-socket argument to kubeadm init. See Installing runtime.
  4. (Optional) Unless otherwise specified, kubeadm uses the network interface associated with the default gateway to set the advertise address for this particular control-plane node's API server. To use a different network interface, specify the --apiserver-advertise-address=<ip-address> argument to kubeadm init. To deploy an IPv6 Kubernetes cluster using IPv6 addressing, you must specify an IPv6 address, for example --apiserver-advertise-address=fd00::101
  5. (Optional) Run kubeadm config images pull prior to kubeadm init to verify connectivity to the gcr.io container image registry.

To initialize the control-plane node run:

kubeadm init <args>

Considerations about apiserver-advertise-address and ControlPlaneEndpoint

While --apiserver-advertise-address can be used to set the advertise address for this particular control-plane node's API server, --control-plane-endpoint can be used to set the shared endpoint for all control-plane nodes.

--control-plane-endpoint allows both IP addresses and DNS names that can map to IP addresses. Please contact your network administrator to evaluate possible solutions with respect to such mapping.

Here is an example mapping: cluster-endpoint

Where is the IP address of this node and cluster-endpoint is a custom DNS name that maps to this IP. This will allow you to pass --control-plane-endpoint=cluster-endpoint to kubeadm init and pass the same DNS name to kubeadm join. Later you can modify cluster-endpoint to point to the address of your load-balancer in an high availability scenario.

Turning a single control plane cluster created without --control-plane-endpoint into a highly available cluster is not supported by kubeadm.

More information

For more information about kubeadm init arguments, see the kubeadm reference guide.

To configure kubeadm init with a configuration file see Using kubeadm init with a configuration file.

To customize control plane components, including optional IPv6 assignment to liveness probe for control plane components and etcd server, provide extra arguments to each component as documented in custom arguments.

To run kubeadm init again, you must first tear down the cluster.

If you join a node with a different architecture to your cluster, make sure that your deployed DaemonSets have container image support for this architecture.

kubeadm init first runs a series of prechecks to ensure that the machine is ready to run Kubernetes. These prechecks expose warnings and exit on errors. kubeadm init then downloads and installs the cluster control plane components. This may take several minutes. After it finishes you should see:

Your Kubernetes control-plane has initialized successfully!

To start using your cluster, you need to run the following as a regular user:

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

You should now deploy a Pod network to the cluster.
Run "kubectl apply -f [podnetwork].yaml" with one of the options listed at:

You can now join any number of machines by running the following on each node
as root:

  kubeadm join <control-plane-host>:<control-plane-port> --token <token> --discovery-token-ca-cert-hash sha256:<hash>

To make kubectl work for your non-root user, run these commands, which are also part of the kubeadm init output:

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

Alternatively, if you are the root user, you can run:

export KUBECONFIG=/etc/kubernetes/admin.conf
Warning: Kubeadm signs the certificate in the admin.conf to have Subject: O = system:masters, CN = kubernetes-admin. system:masters is a break-glass, super user group that bypasses the authorization layer (e.g. RBAC). Do not share the admin.conf file with anyone and instead grant users custom permissions by generating them a kubeconfig file using the kubeadm kubeconfig user command.

Make a record of the kubeadm join command that kubeadm init outputs. You need this command to join nodes to your cluster.

The token is used for mutual authentication between the control-plane node and the joining nodes. The token included here is secret. Keep it safe, because anyone with this token can add authenticated nodes to your cluster. These tokens can be listed, created, and deleted with the kubeadm token command. See the kubeadm reference guide.

Installing a Pod network add-on


This section contains important information about networking setup and deployment order. Read all of this advice carefully before proceeding.

You must deploy a Container Network Interface (CNI) based Pod network add-on so that your Pods can communicate with each other. Cluster DNS (CoreDNS) will not start up before a network is installed.

  • Take care that your Pod network must not overlap with any of the host networks: you are likely to see problems if there is any overlap. (If you find a collision between your network plugin's preferred Pod network and some of your host networks, you should think of a suitable CIDR block to use instead, then use that during kubeadm init with --pod-network-cidr and as a replacement in your network plugin's YAML).

  • By default, kubeadm sets up your cluster to use and enforce use of RBAC (role based access control). Make sure that your Pod network plugin supports RBAC, and so do any manifests that you use to deploy it.

  • If you want to use IPv6--either dual-stack, or single-stack IPv6 only networking--for your cluster, make sure that your Pod network plugin supports IPv6. IPv6 support was added to CNI in v0.6.0.

Note: Kubeadm should be CNI agnostic and the validation of CNI providers is out of the scope of our current e2e testing. If you find an issue related to a CNI plugin you should log a ticket in its respective issue tracker instead of the kubeadm or kubernetes issue trackers.

Several external projects provide Kubernetes Pod networks using CNI, some of which also support Network Policy.

See a list of add-ons that implement the Kubernetes networking model.

You can install a Pod network add-on with the following command on the control-plane node or a node that has the kubeconfig credentials:

kubectl apply -f <add-on.yaml>

You can install only one Pod network per cluster.

Once a Pod network has been installed, you can confirm that it is working by checking that the CoreDNS Pod is Running in the output of kubectl get pods --all-namespaces. And once the CoreDNS Pod is up and running, you can continue by joining your nodes.

If your network is not working or CoreDNS is not in the Running state, check out the troubleshooting guide for kubeadm.

Control plane node isolation

By default, your cluster will not schedule Pods on the control-plane node for security reasons. If you want to be able to schedule Pods on the control-plane node, for example for a single-machine Kubernetes cluster for development, run:

kubectl taint nodes --all node-role.kubernetes.io/master-

With output looking something like:

node "test-01" untainted
taint "node-role.kubernetes.io/master:" not found
taint "node-role.kubernetes.io/master:" not found

This will remove the node-role.kubernetes.io/master taint from any nodes that have it, including the control-plane node, meaning that the scheduler will then be able to schedule Pods everywhere.

Joining your nodes

The nodes are where your workloads (containers and Pods, etc) run. To add new nodes to your cluster do the following for each machine:

  • SSH to the machine
  • Become root (e.g. sudo su -)
  • Run the command that was output by kubeadm init. For example:
kubeadm join --token <token> <control-plane-host>:<control-plane-port> --discovery-token-ca-cert-hash sha256:<hash>

If you do not have the token, you can get it by running the following command on the control-plane node:

kubeadm token list

The output is similar to this:

TOKEN                    TTL  EXPIRES              USAGES           DESCRIPTION            EXTRA GROUPS
8ewj1p.9r9hcjoqgajrj4gi  23h  2018-06-12T02:51:28Z authentication,  The default bootstrap  system:
                                                   signing          token generated by     bootstrappers:
                                                                    'kubeadm init'.        kubeadm:

By default, tokens expire after 24 hours. If you are joining a node to the cluster after the current token has expired, you can create a new token by running the following command on the control-plane node:

kubeadm token create

The output is similar to this:


If you don't have the value of --discovery-token-ca-cert-hash, you can get it by running the following command chain on the control-plane node:

openssl x509 -pubkey -in /etc/kubernetes/pki/ca.crt | openssl rsa -pubin -outform der 2>/dev/null | \
   openssl dgst -sha256 -hex | sed 's/^.* //'

The output is similar to:

Note: To specify an IPv6 tuple for <control-plane-host>:<control-plane-port>, IPv6 address must be enclosed in square brackets, for example: [fd00::101]:2073.

The output should look something like:

[preflight] Running pre-flight checks

... (log output of join workflow) ...

Node join complete:
* Certificate signing request sent to control-plane and response
* Kubelet informed of new secure connection details.

Run 'kubectl get nodes' on control-plane to see this machine join.

A few seconds later, you should notice this node in the output from kubectl get nodes when run on the control-plane node.

(Optional) Controlling your cluster from machines other than the control-plane node

In order to get a kubectl on some other computer (e.g. laptop) to talk to your cluster, you need to copy the administrator kubeconfig file from your control-plane node to your workstation like this:

scp root@<control-plane-host>:/etc/kubernetes/admin.conf .
kubectl --kubeconfig ./admin.conf get nodes

The example above assumes SSH access is enabled for root. If that is not the case, you can copy the admin.conf file to be accessible by some other user and scp using that other user instead.

The admin.conf file gives the user superuser privileges over the cluster. This file should be used sparingly. For normal users, it's recommended to generate an unique credential to which you grant privileges. You can do this with the kubeadm alpha kubeconfig user --client-name <CN> command. That command will print out a KubeConfig file to STDOUT which you should save to a file and distribute to your user. After that, grant privileges by using kubectl create (cluster)rolebinding.

(Optional) Proxying API Server to localhost

If you want to connect to the API Server from outside the cluster you can use kubectl proxy:

scp root@<control-plane-host>:/etc/kubernetes/admin.conf .
kubectl --kubeconfig ./admin.conf proxy

You can now access the API Server locally at http://localhost:8001/api/v1

Clean up

If you used disposable servers for your cluster, for testing, you can switch those off and do no further clean up. You can use kubectl config delete-cluster to delete your local references to the cluster.

However, if you want to deprovision your cluster more cleanly, you should first drain the node and make sure that the node is empty, then deconfigure the node.

Remove the node

Talking to the control-plane node with the appropriate credentials, run:

kubectl drain <node name> --delete-local-data --force --ignore-daemonsets

Before removing the node, reset the state installed by kubeadm:

kubeadm reset

The reset process does not reset or clean up iptables rules or IPVS tables. If you wish to reset iptables, you must do so manually:

iptables -F && iptables -t nat -F && iptables -t mangle -F && iptables -X

If you want to reset the IPVS tables, you must run the following command:

ipvsadm -C

Now remove the node:

kubectl delete node <node name>

If you wish to start over, run kubeadm init or kubeadm join with the appropriate arguments.

Clean up the control plane

You can use kubeadm reset on the control plane host to trigger a best-effort clean up.

See the kubeadm reset reference documentation for more information about this subcommand and its options.

What's next


Version skew policy

The kubeadm tool of version v1.21 may deploy clusters with a control plane of version v1.21 or v1.20. kubeadm v1.21 can also upgrade an existing kubeadm-created cluster of version v1.20.

Due to that we can't see into the future, kubeadm CLI v1.21 may or may not be able to deploy v1.22 clusters.

These resources provide more information on supported version skew between kubelets and the control plane, and other Kubernetes components:


Cluster resilience

The cluster created here has a single control-plane node, with a single etcd database running on it. This means that if the control-plane node fails, your cluster may lose data and may need to be recreated from scratch.


Platform compatibility

kubeadm deb/rpm packages and binaries are built for amd64, arm (32-bit), arm64, ppc64le, and s390x following the multi-platform proposal.

Multiplatform container images for the control plane and addons are also supported since v1.12.

Only some of the network providers offer solutions for all platforms. Please consult the list of network providers above or the documentation from each provider to figure out whether the provider supports your chosen platform.


If you are running into difficulties with kubeadm, please consult our troubleshooting docs.

2.1.4 - Customizing control plane configuration with kubeadm

FEATURE STATE: Kubernetes v1.12 [stable]

The kubeadm ClusterConfiguration object exposes the field extraArgs that can override the default flags passed to control plane components such as the APIServer, ControllerManager and Scheduler. The components are defined using the following fields:

  • apiServer
  • controllerManager
  • scheduler

The extraArgs field consist of key: value pairs. To override a flag for a control plane component:

  1. Add the appropriate fields to your configuration.
  2. Add the flags to override to the field.
  3. Run kubeadm init with --config <YOUR CONFIG YAML>.

For more details on each field in the configuration you can navigate to our API reference pages.

Note: You can generate a ClusterConfiguration object with default values by running kubeadm config print init-defaults and saving the output to a file of your choice.

APIServer flags

For details, see the reference documentation for kube-apiserver.

Example usage:

apiVersion: kubeadm.k8s.io/v1beta2
kind: ClusterConfiguration
kubernetesVersion: v1.16.0
    anonymous-auth: "false"
    enable-admission-plugins: AlwaysPullImages,DefaultStorageClass
    audit-log-path: /home/johndoe/audit.log

ControllerManager flags

For details, see the reference documentation for kube-controller-manager.

Example usage:

apiVersion: kubeadm.k8s.io/v1beta2
kind: ClusterConfiguration
kubernetesVersion: v1.16.0
    cluster-signing-key-file: /home/johndoe/keys/ca.key
    deployment-controller-sync-period: "50"

Scheduler flags

For details, see the reference documentation for kube-scheduler.

Example usage:

apiVersion: kubeadm.k8s.io/v1beta2
kind: ClusterConfiguration
kubernetesVersion: v1.16.0
    config: /home/johndoe/schedconfig.yaml
    kubeconfig: /home/johndoe/kubeconfig.yaml

2.1.5 - Options for Highly Available topology

This page explains the two options for configuring the topology of your highly available (HA) Kubernetes clusters.

You can set up an HA cluster:

  • With stacked control plane nodes, where etcd nodes are colocated with control plane nodes
  • With external etcd nodes, where etcd runs on separate nodes from the control plane

You should carefully consider the advantages and disadvantages of each topology before setting up an HA cluster.

Note: kubeadm bootstraps the etcd cluster statically. Read the etcd Clustering Guide for more details.

Stacked etcd topology

A stacked HA cluster is a topology where the distributed data storage cluster provided by etcd is stacked on top of the cluster formed by the nodes managed by kubeadm that run control plane components.

Each control plane node runs an instance of the kube-apiserver, kube-scheduler, and kube-controller-manager. The kube-apiserver is exposed to worker nodes using a load balancer.

Each control plane node creates a local etcd member and this etcd member communicates only with the kube-apiserver of this node. The same applies to the local kube-controller-manager and kube-scheduler instances.

This topology couples the control planes and etcd members on the same nodes. It is simpler to set up than a cluster with external etcd nodes, and simpler to manage for replication.

However, a stacked cluster runs the risk of failed coupling. If one node goes down, both an etcd member and a control plane instance are lost, and redundancy is compromised. You can mitigate this risk by adding more control plane nodes.

You should therefore run a minimum of three stacked control plane nodes for an HA cluster.

This is the default topology in kubeadm. A local etcd member is created automatically on control plane nodes when using kubeadm init and kubeadm join --control-plane.

Stacked etcd topology

External etcd topology

An HA cluster with external etcd is a topology where the distributed data storage cluster provided by etcd is external to the cluster formed by the nodes that run control plane components.

Like the stacked etcd topology, each control plane node in an external etcd topology runs an instance of the kube-apiserver, kube-scheduler, and kube-controller-manager. And the kube-apiserver is exposed to worker nodes using a load balancer. However, etcd members run on separate hosts, and each etcd host communicates with the kube-apiserver of each control plane node.

This topology decouples the control plane and etcd member. It therefore provides an HA setup where losing a control plane instance or an etcd member has less impact and does not affect the cluster redundancy as much as the stacked HA topology.

However, this topology requires twice the number of hosts as the stacked HA topology. A minimum of three hosts for control plane nodes and three hosts for etcd nodes are required for an HA cluster with this topology.

External etcd topology

What's next

2.1.6 - Creating Highly Available clusters with kubeadm

This page explains two different approaches to setting up a highly available Kubernetes cluster using kubeadm:

  • With stacked control plane nodes. This approach requires less infrastructure. The etcd members and control plane nodes are co-located.
  • With an external etcd cluster. This approach requires more infrastructure. The control plane nodes and etcd members are separated.

Before proceeding, you should carefully consider which approach best meets the needs of your applications and environment. This comparison topic outlines the advantages and disadvantages of each.

If you encounter issues with setting up the HA cluster, please provide us with feedback in the kubeadm issue tracker.

See also The upgrade documentation.

Caution: This page does not address running your cluster on a cloud provider. In a cloud environment, neither approach documented here works with Service objects of type LoadBalancer, or with dynamic PersistentVolumes.

Before you begin

For both methods you need this infrastructure:

  • Three machines that meet kubeadm's minimum requirements for the control-plane nodes
  • Three machines that meet kubeadm's minimum requirements for the workers
  • Full network connectivity between all machines in the cluster (public or private network)
  • sudo privileges on all machines
  • SSH access from one device to all nodes in the system
  • kubeadm and kubelet installed on all machines. kubectl is optional.

For the external etcd cluster only, you also need:

  • Three additional machines for etcd members

First steps for both methods

Create load balancer for kube-apiserver

Note: There are many configurations for load balancers. The following example is only one option. Your cluster requirements may need a different configuration.
  1. Create a kube-apiserver load balancer with a name that resolves to DNS.

    • In a cloud environment you should place your control plane nodes behind a TCP forwarding load balancer. This load balancer distributes traffic to all healthy control plane nodes in its target list. The health check for an apiserver is a TCP check on the port the kube-apiserver listens on (default value :6443).

    • It is not recommended to use an IP address directly in a cloud environment.

    • The load balancer must be able to communicate with all control plane nodes on the apiserver port. It must also allow incoming traffic on its listening port.

    • Make sure the address of the load balancer always matches the address of kubeadm's ControlPlaneEndpoint.

    • Read the Options for Software Load Balancing guide for more details.

  2. Add the first control plane nodes to the load balancer and test the connection:

    • A connection refused error is expected because the apiserver is not yet running. A timeout, however, means the load balancer cannot communicate with the control plane node. If a timeout occurs, reconfigure the load balancer to communicate with the control plane node.
  3. Add the remaining control plane nodes to the load balancer target group.

Stacked control plane and etcd nodes

Steps for the first control plane node

  1. Initialize the control plane:

    sudo kubeadm init --control-plane-endpoint "LOAD_BALANCER_DNS:LOAD_BALANCER_PORT" --upload-certs
    • You can use the --kubernetes-version flag to set the Kubernetes version to use. It is recommended that the versions of kubeadm, kubelet, kubectl and Kubernetes match.

    • The --control-plane-endpoint flag should be set to the address or DNS and port of the load balancer.

    • The --upload-certs flag is used to upload the certificates that should be shared across all the control-plane instances to the cluster. If instead, you prefer to copy certs across control-plane nodes manually or using automation tools, please remove this flag and refer to Manual certificate distribution section below.

    Note: The kubeadm init flags --config and --certificate-key cannot be mixed, therefore if you want to use the kubeadm configuration you must add the certificateKey field in the appropriate config locations (under InitConfiguration and JoinConfiguration: controlPlane).
    Note: Some CNI network plugins require additional configuration, for example specifying the pod IP CIDR, while others do not. See the CNI network documentation. To add a pod CIDR pass the flag --pod-network-cidr, or if you are using a kubeadm configuration file set the podSubnet field under the networking object of ClusterConfiguration.
    • The output looks similar to:

      You can now join any number of control-plane node by running the following command on each as a root:
          kubeadm join --token 9vr73a.a8uxyaju799qwdjv --discovery-token-ca-cert-hash sha256:7c2e69131a36ae2a042a339b33381c6d0d43887e2de83720eff5359e26aec866 --control-plane --certificate-key f8902e114ef118304e561c3ecd4d0b543adc226b7a07f675f56564185ffe0c07
      Please note that the certificate-key gives access to cluster sensitive data, keep it secret!
      As a safeguard, uploaded-certs will be deleted in two hours; If necessary, you can use kubeadm init phase upload-certs to reload certs afterward.
      Then you can join any number of worker nodes by running the following on each as root:
          kubeadm join --token 9vr73a.a8uxyaju799qwdjv --discovery-token-ca-cert-hash sha256:7c2e69131a36ae2a042a339b33381c6d0d43887e2de83720eff5359e26aec866
    • Copy this output to a text file. You will need it later to join control plane and worker nodes to the cluster.

    • When --upload-certs is used with kubeadm init, the certificates of the primary control plane are encrypted and uploaded in the kubeadm-certs Secret.

    • To re-upload the certificates and generate a new decryption key, use the following command on a control plane node that is already joined to the cluster:

      sudo kubeadm init phase upload-certs --upload-certs
    • You can also specify a custom --certificate-key during init that can later be used by join. To generate such a key you can use the following command:

      kubeadm certs certificate-key
    Note: The kubeadm-certs Secret and decryption key expire after two hours.
    Caution: As stated in the command output, the certificate key gives access to cluster sensitive data, keep it secret!
  2. Apply the CNI plugin of your choice: Follow these instructions to install the CNI provider. Make sure the configuration corresponds to the Pod CIDR specified in the kubeadm configuration file if applicable.

    In this example we are using Weave Net:

    kubectl apply -f "https://cloud.weave.works/k8s/net?k8s-version=$(kubectl version | base64 | tr -d '\n')"
  3. Type the following and watch the pods of the control plane components get started:

    kubectl get pod -n kube-system -w

Steps for the rest of the control plane nodes

Note: Since kubeadm version 1.15 you can join multiple control-plane nodes in parallel. Prior to this version, you must join new control plane nodes sequentially, only after the first node has finished initializing.

For each additional control plane node you should:

  1. Execute the join command that was previously given to you by the kubeadm init output on the first node. It should look something like this:

    sudo kubeadm join --token 9vr73a.a8uxyaju799qwdjv --discovery-token-ca-cert-hash sha256:7c2e69131a36ae2a042a339b33381c6d0d43887e2de83720eff5359e26aec866 --control-plane --certificate-key f8902e114ef118304e561c3ecd4d0b543adc226b7a07f675f56564185ffe0c07
    • The --control-plane flag tells kubeadm join to create a new control plane.
    • The --certificate-key ... will cause the control plane certificates to be downloaded from the kubeadm-certs Secret in the cluster and be decrypted using the given key.

External etcd nodes

Setting up a cluster with external etcd nodes is similar to the procedure used for stacked etcd with the exception that you should setup etcd first, and you should pass the etcd information in the kubeadm config file.

Set up the etcd cluster

  1. Follow these instructions to set up the etcd cluster.

  2. Setup SSH as described here.

  3. Copy the following files from any etcd node in the cluster to the first control plane node:

    export CONTROL_PLANE="ubuntu@"
    scp /etc/kubernetes/pki/etcd/ca.crt "${CONTROL_PLANE}":
    scp /etc/kubernetes/pki/apiserver-etcd-client.crt "${CONTROL_PLANE}":
    scp /etc/kubernetes/pki/apiserver-etcd-client.key "${CONTROL_PLANE}":
    • Replace the value of CONTROL_PLANE with the user@host of the first control-plane node.

Set up the first control plane node

  1. Create a file called kubeadm-config.yaml with the following contents:

    apiVersion: kubeadm.k8s.io/v1beta2
    kind: ClusterConfiguration
    kubernetesVersion: stable
            - https://ETCD_0_IP:2379
            - https://ETCD_1_IP:2379
            - https://ETCD_2_IP:2379
            caFile: /etc/kubernetes/pki/etcd/ca.crt
            certFile: /etc/kubernetes/pki/apiserver-etcd-client.crt
            keyFile: /etc/kubernetes/pki/apiserver-etcd-client.key
Note: The difference between stacked etcd and external etcd here is that the external etcd setup requires a configuration file with the etcd endpoints under the external object for etcd. In the case of the stacked etcd topology this is managed automatically.
  • Replace the following variables in the config template with the appropriate values for your cluster:
- `ETCD_0_IP`
- `ETCD_1_IP`
- `ETCD_2_IP`

The following steps are similar to the stacked etcd setup:

  1. Run sudo kubeadm init --config kubeadm-config.yaml --upload-certs on this node.

  2. Write the output join commands that are returned to a text file for later use.

  3. Apply the CNI plugin of your choice. The given example is for Weave Net:

    kubectl apply -f "https://cloud.weave.works/k8s/net?k8s-version=$(kubectl version | base64 | tr -d '\n')"

Steps for the rest of the control plane nodes

The steps are the same as for the stacked etcd setup:

  • Make sure the first control plane node is fully initialized.
  • Join each control plane node with the join command you saved to a text file. It's recommended to join the control plane nodes one at a time.
  • Don't forget that the decryption key from --certificate-key expires after two hours, by default.

Common tasks after bootstrapping control plane

Install workers

Worker nodes can be joined to the cluster with the command you stored previously as the output from the kubeadm init command:

sudo kubeadm join --token 9vr73a.a8uxyaju799qwdjv --discovery-token-ca-cert-hash sha256:7c2e69131a36ae2a042a339b33381c6d0d43887e2de83720eff5359e26aec866

Manual certificate distribution

If you choose to not use kubeadm init with the --upload-certs flag this means that you are going to have to manually copy the certificates from the primary control plane node to the joining control plane nodes.

There are many ways to do this. In the following example we are using ssh and scp:

SSH is required if you want to control all nodes from a single machine.

  1. Enable ssh-agent on your main device that has access to all other nodes in the system:

    eval $(ssh-agent)
  2. Add your SSH identity to the session:

    ssh-add ~/.ssh/path_to_private_key
  3. SSH between nodes to check that the connection is working correctly.

    • When you SSH to any node, make sure to add the -A flag:

      ssh -A
    • When using sudo on any node, make sure to preserve the environment so SSH forwarding works:

      sudo -E -s
  4. After configuring SSH on all the nodes you should run the following script on the first control plane node after running kubeadm init. This script will copy the certificates from the first control plane node to the other control plane nodes:

    In the following example, replace CONTROL_PLANE_IPS with the IP addresses of the other control plane nodes.

    USER=ubuntu # customizable
    for host in ${CONTROL_PLANE_IPS}; do
        scp /etc/kubernetes/pki/ca.crt "${USER}"@$host:
        scp /etc/kubernetes/pki/ca.key "${USER}"@$host:
        scp /etc/kubernetes/pki/sa.key "${USER}"@$host:
        scp /etc/kubernetes/pki/sa.pub "${USER}"@$host:
        scp /etc/kubernetes/pki/front-proxy-ca.crt "${USER}"@$host:
        scp /etc/kubernetes/pki/front-proxy-ca.key "${USER}"@$host:
        scp /etc/kubernetes/pki/etcd/ca.crt "${USER}"@$host:etcd-ca.crt
        # Quote this line if you are using external etcd
        scp /etc/kubernetes/pki/etcd/ca.key "${USER}"@$host:etcd-ca.key
    Caution: Copy only the certificates in the above list. kubeadm will take care of generating the rest of the certificates with the required SANs for the joining control-plane instances. If you copy all the certificates by mistake, the creation of additional nodes could fail due to a lack of required SANs.
  5. Then on each joining control plane node you have to run the following script before running kubeadm join. This script will move the previously copied certificates from the home directory to /etc/kubernetes/pki:

    USER=ubuntu # customizable
    mkdir -p /etc/kubernetes/pki/etcd
    mv /home/${USER}/ca.crt /etc/kubernetes/pki/
    mv /home/${USER}/ca.key /etc/kubernetes/pki/
    mv /home/${USER}/sa.pub /etc/kubernetes/pki/
    mv /home/${USER}/sa.key /etc/kubernetes/pki/
    mv /home/${USER}/front-proxy-ca.crt /etc/kubernetes/pki/
    mv /home/${USER}/front-proxy-ca.key /etc/kubernetes/pki/
    mv /home/${USER}/etcd-ca.crt /etc/kubernetes/pki/etcd/ca.crt
    # Quote this line if you are using external etcd
    mv /home/${USER}/etcd-ca.key /etc/kubernetes/pki/etcd/ca.key

2.1.7 - Set up a High Availability etcd cluster with kubeadm

Note: While kubeadm is being used as the management tool for external etcd nodes in this guide, please note that kubeadm does not plan to support certificate rotation or upgrades for such nodes. The long term plan is to empower the tool etcdadm to manage these aspects.

Kubeadm defaults to running a single member etcd cluster in a static pod managed by the kubelet on the control plane node. This is not a high availability setup as the etcd cluster contains only one member and cannot sustain any members becoming unavailable. This task walks through the process of creating a high availability etcd cluster of three members that can be used as an external etcd when using kubeadm to set up a kubernetes cluster.

Before you begin

  • Three hosts that can talk to each other over ports 2379 and 2380. This document assumes these default ports. However, they are configurable through the kubeadm config file.
  • Each host must have docker, kubelet, and kubeadm installed.
  • Each host should have access to the Kubernetes container image registry (k8s.gcr.io) or list/pull the required etcd image using kubeadm config images list/pull. This guide will setup etcd instances as static pods managed by a kubelet.
  • Some infrastructure to copy files between hosts. For example ssh and scp can satisfy this requirement.

Setting up the cluster

The general approach is to generate all certs on one node and only distribute the necessary files to the other nodes.

Note: kubeadm contains all the necessary crytographic machinery to generate the certificates described below; no other cryptographic tooling is required for this example.
  1. Configure the kubelet to be a service manager for etcd.

    Note: You must do this on every host where etcd should be running.
    Since etcd was created first, you must override the service priority by creating a new unit file that has higher precedence than the kubeadm-provided kubelet unit file.

    cat << EOF > /etc/systemd/system/kubelet.service.d/20-etcd-service-manager.conf
    #  Replace "systemd" with the cgroup driver of your container runtime. The default value in the kubelet is "cgroupfs".
    ExecStart=/usr/bin/kubelet --address= --pod-manifest-path=/etc/kubernetes/manifests --cgroup-driver=systemd
    systemctl daemon-reload
    systemctl restart kubelet

    Check the kubelet status to ensure it is running.

    systemctl status kubelet
  2. Create configuration files for kubeadm.

    Generate one kubeadm configuration file for each host that will have an etcd member running on it using the following script.

    # Update HOST0, HOST1, and HOST2 with the IPs or resolvable names of your hosts
    export HOST0=
    export HOST1=
    export HOST2=
    # Create temp directories to store files that will end up on other hosts.
    mkdir -p /tmp/${HOST0}/ /tmp/${HOST1}/ /tmp/${HOST2}/
    ETCDHOSTS=(${HOST0} ${HOST1} ${HOST2})
    NAMES=("infra0" "infra1" "infra2")
    for i in "${!ETCDHOSTS[@]}"; do
    cat << EOF > /tmp/${HOST}/kubeadmcfg.yaml
    apiVersion: "kubeadm.k8s.io/v1beta2"
    kind: ClusterConfiguration
            - "${HOST}"
            - "${HOST}"
                initial-cluster: ${NAMES[0]}=https://${ETCDHOSTS[0]}:2380,${NAMES[1]}=https://${ETCDHOSTS[1]}:2380,${NAMES[2]}=https://${ETCDHOSTS[2]}:2380
                initial-cluster-state: new
                name: ${NAME}
                listen-peer-urls: https://${HOST}:2380
                listen-client-urls: https://${HOST}:2379
                advertise-client-urls: https://${HOST}:2379
                initial-advertise-peer-urls: https://${HOST}:2380
  3. Generate the certificate authority

    If you already have a CA then the only action that is copying the CA's crt and key file to /etc/kubernetes/pki/etcd/ca.crt and /etc/kubernetes/pki/etcd/ca.key. After those files have been copied, proceed to the next step, "Create certificates for each member".

    If you do not already have a CA then run this command on $HOST0 (where you generated the configuration files for kubeadm).

    kubeadm init phase certs etcd-ca

    This creates two files

    • /etc/kubernetes/pki/etcd/ca.crt
    • /etc/kubernetes/pki/etcd/ca.key
  4. Create certificates for each member

    kubeadm init phase certs etcd-server --config=/tmp/${HOST2}/kubeadmcfg.yaml
    kubeadm init phase certs etcd-peer --config=/tmp/${HOST2}/kubeadmcfg.yaml
    kubeadm init phase certs etcd-healthcheck-client --config=/tmp/${HOST2}/kubeadmcfg.yaml
    kubeadm init phase certs apiserver-etcd-client --config=/tmp/${HOST2}/kubeadmcfg.yaml
    cp -R /etc/kubernetes/pki /tmp/${HOST2}/
    # cleanup non-reusable certificates
    find /etc/kubernetes/pki -not -name ca.crt -not -name ca.key -type f -delete
    kubeadm init phase certs etcd-server --config=/tmp/${HOST1}/kubeadmcfg.yaml
    kubeadm init phase certs etcd-peer --config=/tmp/${HOST1}/kubeadmcfg.yaml
    kubeadm init phase certs etcd-healthcheck-client --config=/tmp/${HOST1}/kubeadmcfg.yaml
    kubeadm init phase certs apiserver-etcd-client --config=/tmp/${HOST1}/kubeadmcfg.yaml
    cp -R /etc/kubernetes/pki /tmp/${HOST1}/
    find /etc/kubernetes/pki -not -name ca.crt -not -name ca.key -type f -delete
    kubeadm init phase certs etcd-server --config=/tmp/${HOST0}/kubeadmcfg.yaml
    kubeadm init phase certs etcd-peer --config=/tmp/${HOST0}/kubeadmcfg.yaml
    kubeadm init phase certs etcd-healthcheck-client --config=/tmp/${HOST0}/kubeadmcfg.yaml
    kubeadm init phase certs apiserver-etcd-client --config=/tmp/${HOST0}/kubeadmcfg.yaml
    # No need to move the certs because they are for HOST0
    # clean up certs that should not be copied off this host
    find /tmp/${HOST2} -name ca.key -type f -delete
    find /tmp/${HOST1} -name ca.key -type f -delete
  5. Copy certificates and kubeadm configs

    The certificates have been generated and now they must be moved to their respective hosts.

    scp -r /tmp/${HOST}/* ${USER}@${HOST}:
    ssh ${USER}@${HOST}
    USER@HOST $ sudo -Es
    root@HOST $ chown -R root:root pki
    root@HOST $ mv pki /etc/kubernetes/
  6. Ensure all expected files exist

    The complete list of required files on $HOST0 is:

    └── kubeadmcfg.yaml
    ├── apiserver-etcd-client.crt
    ├── apiserver-etcd-client.key
    └── etcd
        ├── ca.crt
        ├── ca.key
        ├── healthcheck-client.crt
        ├── healthcheck-client.key
        ├── peer.crt
        ├── peer.key
        ├── server.crt
        └── server.key

    On $HOST1:

    └── kubeadmcfg.yaml
    ├── apiserver-etcd-client.crt
    ├── apiserver-etcd-client.key
    └── etcd
        ├── ca.crt
        ├── healthcheck-client.crt
        ├── healthcheck-client.key
        ├── peer.crt
        ├── peer.key
        ├── server.crt
        └── server.key

    On $HOST2

    └── kubeadmcfg.yaml
    ├── apiserver-etcd-client.crt
    ├── apiserver-etcd-client.key
    └── etcd
        ├── ca.crt
        ├── healthcheck-client.crt
        ├── healthcheck-client.key
        ├── peer.crt
        ├── peer.key
        ├── server.crt
        └── server.key
  7. Create the static pod manifests

    Now that the certificates and configs are in place it's time to create the manifests. On each host run the kubeadm command to generate a static manifest for etcd.

    root@HOST0 $ kubeadm init phase etcd local --config=/tmp/${HOST0}/kubeadmcfg.yaml
    root@HOST1 $ kubeadm init phase etcd local --config=/home/ubuntu/kubeadmcfg.yaml
    root@HOST2 $ kubeadm init phase etcd local --config=/home/ubuntu/kubeadmcfg.yaml
  8. Optional: Check the cluster health

    docker run --rm -it \
    --net host \
    -v /etc/kubernetes:/etc/kubernetes k8s.gcr.io/etcd:${ETCD_TAG} etcdctl \
    --cert /etc/kubernetes/pki/etcd/peer.crt \
    --key /etc/kubernetes/pki/etcd/peer.key \
    --cacert /etc/kubernetes/pki/etcd/ca.crt \
    --endpoints https://${HOST0}:2379 endpoint health --cluster
    https://[HOST0 IP]:2379 is healthy: successfully committed proposal: took = 16.283339ms
    https://[HOST1 IP]:2379 is healthy: successfully committed proposal: took = 19.44402ms
    https://[HOST2 IP]:2379 is healthy: successfully committed proposal: took = 35.926451ms
    • Set ${ETCD_TAG} to the version tag of your etcd image. For example 3.4.3-0. To see the etcd image and tag that kubeadm uses execute kubeadm config images list --kubernetes-version ${K8S_VERSION}, where ${K8S_VERSION} is for example v1.17.0
    • Set ${HOST0}to the IP address of the host you are testing.

What's next

Once you have a working 3 member etcd cluster, you can continue setting up a highly available control plane using the external etcd method with kubeadm.

2.1.8 - Configuring each kubelet in your cluster using kubeadm

FEATURE STATE: Kubernetes v1.11 [stable]

The lifecycle of the kubeadm CLI tool is decoupled from the kubelet, which is a daemon that runs on each node within the Kubernetes cluster. The kubeadm CLI tool is executed by the user when Kubernetes is initialized or upgraded, whereas the kubelet is always running in the background.

Since the kubelet is a daemon, it needs to be maintained by some kind of an init system or service manager. When the kubelet is installed using DEBs or RPMs, systemd is configured to manage the kubelet. You can use a different service manager instead, but you need to configure it manually.

Some kubelet configuration details need to be the same across all kubelets involved in the cluster, while other configuration aspects need to be set on a per-kubelet basis to accommodate the different characteristics of a given machine (such as OS, storage, and networking). You can manage the configuration of your kubelets manually, but kubeadm now provides a KubeletConfiguration API type for managing your kubelet configurations centrally.

Kubelet configuration patterns

The following sections describe patterns to kubelet configuration that are simplified by using kubeadm, rather than managing the kubelet configuration for each Node manually.

Propagating cluster-level configuration to each kubelet

You can provide the kubelet with default values to be used by kubeadm init and kubeadm join commands. Interesting examples include using a different CRI runtime or setting the default subnet used by services.

If you want your services to use the subnet as the default for services, you can pass the --service-cidr parameter to kubeadm:

kubeadm init --service-cidr

Virtual IPs for services are now allocated from this subnet. You also need to set the DNS address used by the kubelet, using the --cluster-dns flag. This setting needs to be the same for every kubelet on every manager and Node in the cluster. The kubelet provides a versioned, structured API object that can configure most parameters in the kubelet and push out this configuration to each running kubelet in the cluster. This object is called KubeletConfiguration. The KubeletConfiguration allows the user to specify flags such as the cluster DNS IP addresses expressed as a list of values to a camelCased key, illustrated by the following example:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration

For more details on the KubeletConfiguration have a look at this section.

Providing instance-specific configuration details

Some hosts require specific kubelet configurations due to differences in hardware, operating system, networking, or other host-specific parameters. The following list provides a few examples.

  • The path to the DNS resolution file, as specified by the --resolv-conf kubelet configuration flag, may differ among operating systems, or depending on whether you are using systemd-resolved. If this path is wrong, DNS resolution will fail on the Node whose kubelet is configured incorrectly.

  • The Node API object .metadata.name is set to the machine's hostname by default, unless you are using a cloud provider. You can use the --hostname-override flag to override the default behavior if you need to specify a Node name different from the machine's hostname.

  • Currently, the kubelet cannot automatically detect the cgroup driver used by the CRI runtime, but the value of --cgroup-driver must match the cgroup driver used by the CRI runtime to ensure the health of the kubelet.

  • Depending on the CRI runtime your cluster uses, you may need to specify different flags to the kubelet. For instance, when using Docker, you need to specify flags such as --network-plugin=cni, but if you are using an external runtime, you need to specify --container-runtime=remote and specify the CRI endpoint using the --container-runtime-endpoint=<path>.

You can specify these flags by configuring an individual kubelet's configuration in your service manager, such as systemd.

Configure kubelets using kubeadm

It is possible to configure the kubelet that kubeadm will start if a custom KubeletConfiguration API object is passed with a configuration file like so kubeadm ... --config some-config-file.yaml.

By calling kubeadm config print init-defaults --component-configs KubeletConfiguration you can see all the default values for this structure.

Also have a look at the reference for the KubeletConfiguration for more information on the individual fields.

Workflow when using kubeadm init

When you call kubeadm init, the kubelet configuration is marshalled to disk at /var/lib/kubelet/config.yaml, and also uploaded to a ConfigMap in the cluster. The ConfigMap is named kubelet-config-1.X, where X is the minor version of the Kubernetes version you are initializing. A kubelet configuration file is also written to /etc/kubernetes/kubelet.conf with the baseline cluster-wide configuration for all kubelets in the cluster. This configuration file points to the client certificates that allow the kubelet to communicate with the API server. This addresses the need to propagate cluster-level configuration to each kubelet.

To address the second pattern of providing instance-specific configuration details, kubeadm writes an environment file to /var/lib/kubelet/kubeadm-flags.env, which contains a list of flags to pass to the kubelet when it starts. The flags are presented in the file like this:

KUBELET_KUBEADM_ARGS="--flag1=value1 --flag2=value2 ..."

In addition to the flags used when starting the kubelet, the file also contains dynamic parameters such as the cgroup driver and whether to use a different CRI runtime socket (--cri-socket).

After marshalling these two files to disk, kubeadm attempts to run the following two commands, if you are using systemd:

systemctl daemon-reload && systemctl restart kubelet

If the reload and restart are successful, the normal kubeadm init workflow continues.

Workflow when using kubeadm join

When you run kubeadm join, kubeadm uses the Bootstrap Token credential to perform a TLS bootstrap, which fetches the credential needed to download the kubelet-config-1.X ConfigMap and writes it to /var/lib/kubelet/config.yaml. The dynamic environment file is generated in exactly the same way as kubeadm init.

Next, kubeadm runs the following two commands to load the new configuration into the kubelet:

systemctl daemon-reload && systemctl restart kubelet

After the kubelet loads the new configuration, kubeadm writes the /etc/kubernetes/bootstrap-kubelet.conf KubeConfig file, which contains a CA certificate and Bootstrap Token. These are used by the kubelet to perform the TLS Bootstrap and obtain a unique credential, which is stored in /etc/kubernetes/kubelet.conf. When this file is written, the kubelet has finished performing the TLS Bootstrap.

The kubelet drop-in file for systemd

kubeadm ships with configuration for how systemd should run the kubelet. Note that the kubeadm CLI command never touches this drop-in file.

This configuration file installed by the kubeadm DEB or RPM package is written to /etc/systemd/system/kubelet.service.d/10-kubeadm.conf and is used by systemd. It augments the basic kubelet.service for RPM or kubelet.service for DEB:

# This is a file that "kubeadm init" and "kubeadm join" generate at runtime, populating
the KUBELET_KUBEADM_ARGS variable dynamically
# This is a file that the user can use for overrides of the kubelet args as a last resort. Preferably,
# the user should use the .NodeRegistration.KubeletExtraArgs object in the configuration files instead.
# KUBELET_EXTRA_ARGS should be sourced from this file.

This file specifies the default locations for all of the files managed by kubeadm for the kubelet.

  • The KubeConfig file to use for the TLS Bootstrap is /etc/kubernetes/bootstrap-kubelet.conf, but it is only used if /etc/kubernetes/kubelet.conf does not exist.
  • The KubeConfig file with the unique kubelet identity is /etc/kubernetes/kubelet.conf.
  • The file containing the kubelet's ComponentConfig is /var/lib/kubelet/config.yaml.
  • The dynamic environment file that contains KUBELET_KUBEADM_ARGS is sourced from /var/lib/kubelet/kubeadm-flags.env.
  • The file that can contain user-specified flag overrides with KUBELET_EXTRA_ARGS is sourced from /etc/default/kubelet (for DEBs), or /etc/sysconfig/kubelet (for RPMs). KUBELET_EXTRA_ARGS is last in the flag chain and has the highest priority in the event of conflicting settings.

Kubernetes binaries and package contents

The DEB and RPM packages shipped with the Kubernetes releases are:

Package nameDescription
kubeadmInstalls the /usr/bin/kubeadm CLI tool and the kubelet drop-in file for the kubelet.
kubeletInstalls the kubelet binary in /usr/bin and CNI binaries in /opt/cni/bin.
kubectlInstalls the /usr/bin/kubectl binary.
cri-toolsInstalls the /usr/bin/crictl binary from the cri-tools git repository.

2.1.9 - Dual-stack support with kubeadm

FEATURE STATE: Kubernetes v1.21 [beta]

Your Kubernetes cluster can run in dual-stack networking mode, which means that cluster networking lets you use either address family. In a dual-stack cluster, the control plane can assign both an IPv4 address and an IPv6 address to a single Pod or a Service.

Before you begin

You need to have installed the kubeadm tool, following the steps from Installing kubeadm.

For each server that you want to use as a node, make sure it allows IPv6 forwarding. On Linux, you can set this by running run sysctl -w net.ipv6.conf.all.forwarding=1 as the root user on each server.

You need to have an IPv4 and and IPv6 address range to use. Cluster operators typically use private address ranges for IPv4. For IPv6, a cluster operator typically chooses a global unicast address block from within 2000::/3, using a range that is assigned to the operator.
You don't have to route the cluster's IP address ranges to the public internet.

The size of the IP address allocations should be suitable for the number of Pods and Services that you are planning to run.

Note: If you are upgrading an existing cluster then, by default, the kubeadm upgrade command changes the feature gate IPv6DualStack to true if that is not already enabled.
However, kubeadm does not support making modifications to the pod IP address range (“cluster CIDR”) nor to the cluster's Service address range (“Service CIDR”).

Create a dual-stack cluster

To create a dual-stack cluster with kubeadm init you can pass command line arguments similar to the following example:

# These address ranges are examples
kubeadm init --pod-network-cidr=,2001:db8:42:0::/56 --service-cidr=,2001:db8:42:1::/112

To make things clearer, here is an example kubeadm configuration file kubeadm-config.yaml for the primary dual-stack control plane node.

apiVersion: kubeadm.k8s.io/v1beta2
kind: ClusterConfiguration
  IPv6DualStack: true
apiVersion: kubeadm.k8s.io/v1beta2
kind: InitConfiguration
  advertiseAddress: ""
  bindPort: 6443

advertiseAddress in InitConfiguration specifies the IP address that the API Server will advertise it is listening on. The value of advertiseAddress equals the --apiserver-advertise-address flag of kubeadm init

Run kubeadm to initiate the dual-stack control plane node:

kubeadm init --config=kubeadm-config.yaml

Currently, the kube-controller-manager flags --node-cidr-mask-size-ipv4|--node-cidr-mask-size-ipv6 are being left with default values. See enable IPv4/IPv6 dual stack.

Note: The --apiserver-advertise-address flag does not support dual-stack.

Join a node to dual-stack cluster

Before joining a node, make sure that the node has IPv6 routable network interface and allows IPv6 forwarding.

Here is an example kubeadm configuration file kubeadm-config.yaml for joining a worker node to the cluster.

apiVersion: kubeadm.k8s.io/v1beta2
kind: JoinConfiguration

Also, here is an example kubeadm configuration file kubeadm-config.yaml for joining another control plane node to the cluster.

apiVersion: kubeadm.k8s.io/v1beta2
kind: JoinConfiguration
    advertiseAddress: ""
    bindPort: 6443

advertiseAddress in JoinConfiguration.controlPlane specifies the IP address that the API Server will advertise it is listening on. The value of advertiseAddress equals the --apiserver-advertise-address flag of kubeadm join.

kubeadm join --config=kubeadm-config.yaml ...

Create a single-stack cluster

Note: Enabling the dual-stack feature doesn't mean that you need to use dual-stack addressing.
You can deploy a single-stack cluster that has the dual-stack networking feature enabled.

In 1.21 the IPv6DualStack feature is Beta and the feature gate is defaulted to true. To disable the feature you must configure the feature gate to false. Note that once the feature is GA, the feature gate will be removed.

kubeadm init --feature-gates IPv6DualStack=false

To make things more clear, here is an example kubeadm configuration file kubeadm-config.yaml for the single-stack control plane node.

apiVersion: kubeadm.k8s.io/v1beta2
kind: ClusterConfiguration
  IPv6DualStack: false

What's next

2.2 - Installing Kubernetes with kops

This quickstart shows you how to easily install a Kubernetes cluster on AWS. It uses a tool called kops.

kops is an automated provisioning system:

  • Fully automated installation
  • Uses DNS to identify clusters
  • Self-healing: everything runs in Auto-Scaling Groups
  • Multiple OS support (Debian, Ubuntu 16.04 supported, CentOS & RHEL, Amazon Linux and CoreOS) - see the images.md
  • High-Availability support - see the high_availability.md
  • Can directly provision, or generate terraform manifests - see the terraform.md

Before you begin

Creating a cluster

(1/5) Install kops


Download kops from the releases page (it is also convenient to build from source):

Download the latest release with the command:

curl -LO https://github.com/kubernetes/kops/releases/download/$(curl -s https://api.github.com/repos/kubernetes/kops/releases/latest | grep tag_name | cut -d '"' -f 4)/kops-darwin-amd64

To download a specific version, replace the following portion of the command with the specific kops version.

$(curl -s https://api.github.com/repos/kubernetes/kops/releases/latest | grep tag_name | cut -d '"' -f 4)

For example, to download kops version v1.20.0 type:

curl -LO https://github.com/kubernetes/kops/releases/download/v1.20.0/kops-darwin-amd64

Make the kops binary executable.

chmod +x kops-darwin-amd64

Move the kops binary in to your PATH.

sudo mv kops-darwin-amd64 /usr/local/bin/kops

You can also install kops using Homebrew.

brew update && brew install kops

Download the latest release with the command:

curl -LO https://github.com/kubernetes/kops/releases/download/$(curl -s https://api.github.com/repos/kubernetes/kops/releases/latest | grep tag_name | cut -d '"' -f 4)/kops-linux-amd64

To download a specific version of kops, replace the following portion of the command with the specific kops version.

$(curl -s https://api.github.com/repos/kubernetes/kops/releases/latest | grep tag_name | cut -d '"' -f 4)

For example, to download kops version v1.20.0 type:

curl -LO https://github.com/kubernetes/kops/releases/download/v1.20.0/kops-linux-amd64

Make the kops binary executable

chmod +x kops-linux-amd64

Move the kops binary in to your PATH.

sudo mv kops-linux-amd64 /usr/local/bin/kops

You can also install kops using Homebrew.

brew update && brew install kops

(2/5) Create a route53 domain for your cluster

kops uses DNS for discovery, both inside the cluster and outside, so that you can reach the kubernetes API server from clients.

kops has a strong opinion on the cluster name: it should be a valid DNS name. By doing so you will no longer get your clusters confused, you can share clusters with your colleagues unambiguously, and you can reach them without relying on remembering an IP address.

You can, and probably should, use subdomains to divide your clusters. As our example we will use useast1.dev.example.com. The API server endpoint will then be api.useast1.dev.example.com.

A Route53 hosted zone can serve subdomains. Your hosted zone could be useast1.dev.example.com, but also dev.example.com or even example.com. kops works with any of these, so typically you choose for organization reasons (e.g. you are allowed to create records under dev.example.com, but not under example.com).

Let's assume you're using dev.example.com as your hosted zone. You create that hosted zone using the normal process, or with a command such as aws route53 create-hosted-zone --name dev.example.com --caller-reference 1.

You must then set up your NS records in the parent domain, so that records in the domain will resolve. Here, you would create NS records in example.com for dev. If it is a root domain name you would configure the NS records at your domain registrar (e.g. example.com would need to be configured where you bought example.com).

Verify your route53 domain setup (it is the #1 cause of problems!). You can double-check that your cluster is configured correctly if you have the dig tool by running:

dig NS dev.example.com

You should see the 4 NS records that Route53 assigned your hosted zone.

(3/5) Create an S3 bucket to store your clusters state

kops lets you manage your clusters even after installation. To do this, it must keep track of the clusters that you have created, along with their configuration, the keys they are using etc. This information is stored in an S3 bucket. S3 permissions are used to control access to the bucket.

Multiple clusters can use the same S3 bucket, and you can share an S3 bucket between your colleagues that administer the same clusters - this is much easier than passing around kubecfg files. But anyone with access to the S3 bucket will have administrative access to all your clusters, so you don't want to share it beyond the operations team.

So typically you have one S3 bucket for each ops team (and often the name will correspond to the name of the hosted zone above!)

In our example, we chose dev.example.com as our hosted zone, so let's pick clusters.dev.example.com as the S3 bucket name.

  • Export AWS_PROFILE (if you need to select a profile for the AWS CLI to work)

  • Create the S3 bucket using aws s3 mb s3://clusters.dev.example.com

  • You can export KOPS_STATE_STORE=s3://clusters.dev.example.com and then kops will use this location by default. We suggest putting this in your bash profile or similar.

(4/5) Build your cluster configuration

Run kops create cluster to create your cluster configuration:

kops create cluster --zones=us-east-1c useast1.dev.example.com

kops will create the configuration for your cluster. Note that it only creates the configuration, it does not actually create the cloud resources - you'll do that in the next step with a kops update cluster. This give you an opportunity to review the configuration or change it.

It prints commands you can use to explore further:

  • List your clusters with: kops get cluster
  • Edit this cluster with: kops edit cluster useast1.dev.example.com
  • Edit your node instance group: kops edit ig --name=useast1.dev.example.com nodes
  • Edit your master instance group: kops edit ig --name=useast1.dev.example.com master-us-east-1c

If this is your first time using kops, do spend a few minutes to try those out! An instance group is a set of instances, which will be registered as kubernetes nodes. On AWS this is implemented via auto-scaling-groups. You can have several instance groups, for example if you wanted nodes that are a mix of spot and on-demand instances, or GPU and non-GPU instances.

(5/5) Create the cluster in AWS

Run "kops update cluster" to create your cluster in AWS:

kops update cluster useast1.dev.example.com --yes

That takes a few seconds to run, but then your cluster will likely take a few minutes to actually be ready. kops update cluster will be the tool you'll use whenever you change the configuration of your cluster; it applies the changes you have made to the configuration to your cluster - reconfiguring AWS or kubernetes as needed.

For example, after you kops edit ig nodes, then kops update cluster --yes to apply your configuration, and sometimes you will also have to kops rolling-update cluster to roll out the configuration immediately.

Without --yes, kops update cluster will show you a preview of what it is going to do. This is handy for production clusters!

Explore other add-ons

See the list of add-ons to explore other add-ons, including tools for logging, monitoring, network policy, visualization, and control of your Kubernetes cluster.


  • To delete your cluster: kops delete cluster useast1.dev.example.com --yes

What's next

2.3 - Installing Kubernetes with Kubespray

This quickstart helps to install a Kubernetes cluster hosted on GCE, Azure, OpenStack, AWS, vSphere, Packet (bare metal), Oracle Cloud Infrastructure (Experimental) or Baremetal with Kubespray.

Kubespray is a composition of Ansible playbooks, inventory, provisioning tools, and domain knowledge for generic OS/Kubernetes clusters configuration management tasks. Kubespray provides:

  • a highly available cluster
  • composable attributes
  • support for most popular Linux distributions
    • Ubuntu 16.04, 18.04, 20.04
    • CentOS/RHEL/Oracle Linux 7, 8
    • Debian Buster, Jessie, Stretch, Wheezy
    • Fedora 31, 32
    • Fedora CoreOS
    • openSUSE Leap 15
    • Flatcar Container Linux by Kinvolk
  • continuous integration tests

To choose a tool which best fits your use case, read this comparison to kubeadm and kops.

Creating a cluster

(1/5) Meet the underlay requirements

Provision servers with the following requirements:

  • Ansible v2.9 and python-netaddr is installed on the machine that will run Ansible commands
  • Jinja 2.11 (or newer) is required to run the Ansible Playbooks
  • The target servers must have access to the Internet in order to pull docker images. Otherwise, additional configuration is required (See Offline Environment)
  • The target servers are configured to allow IPv4 forwarding
  • Your ssh key must be copied to all the servers part of your inventory
  • The firewalls are not managed, you'll need to implement your own rules the way you used to. in order to avoid any issue during deployment you should disable your firewall
  • If kubespray is ran from non-root user account, correct privilege escalation method should be configured in the target servers. Then the ansible_become flag or command parameters --become or -b should be specified

Kubespray provides the following utilities to help provision your environment:

(2/5) Compose an inventory file

After you provision your servers, create an inventory file for Ansible. You can do this manually or via a dynamic inventory script. For more information, see "Building your own inventory".

(3/5) Plan your cluster deployment

Kubespray provides the ability to customize many aspects of the deployment:

  • Choice deployment mode: kubeadm or non-kubeadm
  • CNI (networking) plugins
  • DNS configuration
  • Choice of control plane: native/binary or containerized
  • Component versions
  • Calico route reflectors
  • Component runtime options
  • Certificate generation methods

Kubespray customizations can be made to a variable file. If you are getting started with Kubespray, consider using the Kubespray defaults to deploy your cluster and explore Kubernetes.

(4/5) Deploy a Cluster

Next, deploy your cluster:

Cluster deployment using ansible-playbook.

ansible-playbook -i your/inventory/inventory.ini cluster.yml -b -v \

Large deployments (100+ nodes) may require specific adjustments for best results.

(5/5) Verify the deployment

Kubespray provides a way to verify inter-pod connectivity and DNS resolve with Netchecker. Netchecker ensures the netchecker-agents pods can resolve DNS requests and ping each over within the default namespace. Those pods mimic similar behavior of the rest of the workloads and serve as cluster health indicators.

Cluster operations

Kubespray provides additional playbooks to manage your cluster: scale and upgrade.

Scale your cluster

You can add worker nodes from your cluster by running the scale playbook. For more information, see "Adding nodes". You can remove worker nodes from your cluster by running the remove-node playbook. For more information, see "Remove nodes".

Upgrade your cluster

You can upgrade your cluster by running the upgrade-cluster playbook. For more information, see "Upgrades".


You can reset your nodes and wipe out all components installed with Kubespray via the reset playbook.

Caution: When running the reset playbook, be sure not to accidentally target your production cluster!


What's next

Check out planned work on Kubespray's roadmap.

3 - Turnkey Cloud Solutions

This page provides a list of Kubernetes certified solution providers. From each provider page, you can learn how to install and setup production ready clusters.

4 - Windows in Kubernetes

4.1 - Intro to Windows support in Kubernetes

Windows applications constitute a large portion of the services and applications that run in many organizations. Windows containers provide a modern way to encapsulate processes and package dependencies, making it easier to use DevOps practices and follow cloud native patterns for Windows applications. Kubernetes has become the defacto standard container orchestrator, and the release of Kubernetes 1.14 includes production support for scheduling Windows containers on Windows nodes in a Kubernetes cluster, enabling a vast ecosystem of Windows applications to leverage the power of Kubernetes. Organizations with investments in Windows-based applications and Linux-based applications don't have to look for separate orchestrators to manage their workloads, leading to increased operational efficiencies across their deployments, regardless of operating system.

Windows containers in Kubernetes

To enable the orchestration of Windows containers in Kubernetes, include Windows nodes in your existing Linux cluster. Scheduling Windows containers in Pods on Kubernetes is similar to scheduling Linux-based containers.

In order to run Windows containers, your Kubernetes cluster must include multiple operating systems, with control plane nodes running Linux and workers running either Windows or Linux depending on your workload needs. Windows Server 2019 is the only Windows operating system supported, enabling Kubernetes Node on Windows (including kubelet, container runtime, and kube-proxy). For a detailed explanation of Windows distribution channels see the Microsoft documentation.

Note: The Kubernetes control plane, including the master components, continues to run on Linux. There are no plans to have a Windows-only Kubernetes cluster.
Note: In this document, when we talk about Windows containers we mean Windows containers with process isolation. Windows containers with Hyper-V isolation is planned for a future release.

Supported Functionality and Limitations

Supported Functionality

Windows OS Version Support

Refer to the following table for Windows operating system support in Kubernetes. A single heterogeneous Kubernetes cluster can have both Windows and Linux worker nodes. Windows containers have to be scheduled on Windows nodes and Linux containers on Linux nodes.

Kubernetes versionWindows Server LTSC releasesWindows Server SAC releases
Kubernetes v1.17Windows Server 2019Windows Server ver 1809
Kubernetes v1.18Windows Server 2019Windows Server ver 1809, Windows Server ver 1903, Windows Server ver 1909
Kubernetes v1.19Windows Server 2019Windows Server ver 1909, Windows Server ver 2004
Kubernetes v1.20Windows Server 2019Windows Server ver 1909, Windows Server ver 2004

Note: Information on the different Windows Server servicing channels including their support models can be found at Windows Server servicing channels.
Note: We don't expect all Windows customers to update the operating system for their apps frequently. Upgrading your applications is what dictates and necessitates upgrading or introducing new nodes to the cluster. For the customers that chose to upgrade their operating system for containers running on Kubernetes, we will offer guidance and step-by-step instructions when we add support for a new operating system version. This guidance will include recommended upgrade procedures for upgrading user applications together with cluster nodes. Windows nodes adhere to Kubernetes version-skew policy (node to control plane versioning) the same way as Linux nodes do today.
Note: The Windows Server Host Operating System is subject to the Windows Server licensing. The Windows Container images are subject to the Supplemental License Terms for Windows containers.
Note: Windows containers with process isolation have strict compatibility rules, where the host OS version must match the container base image OS version. Once we support Windows containers with Hyper-V isolation in Kubernetes, the limitation and compatibility rules will change.

Pause Image

Microsoft maintains a Windows pause infrastructure container at mcr.microsoft.com/oss/kubernetes/pause:1.4.1.


From an API and kubectl perspective, Windows containers behave in much the same way as Linux-based containers. However, there are some notable differences in key functionality which are outlined in the limitation section.

Key Kubernetes elements work the same way in Windows as they do in Linux. In this section, we talk about some of the key workload enablers and how they map to Windows.

  • Pods

    A Pod is the basic building block of Kubernetes–the smallest and simplest unit in the Kubernetes object model that you create or deploy. You may not deploy Windows and Linux containers in the same Pod. All containers in a Pod are scheduled onto a single Node where each Node represents a specific platform and architecture. The following Pod capabilities, properties and events are supported with Windows containers:

    • Single or multiple containers per Pod with process isolation and volume sharing
    • Pod status fields
    • Readiness and Liveness probes
    • postStart & preStop container lifecycle events
    • ConfigMap, Secrets: as environment variables or volumes
    • EmptyDir
    • Named pipe host mounts
    • Resource limits
  • Controllers

    Kubernetes controllers handle the desired state of Pods. The following workload controllers are supported with Windows containers:

    • ReplicaSet
    • ReplicationController
    • Deployments
    • StatefulSets
    • DaemonSet
    • Job
    • CronJob
  • Services

    A Kubernetes Service is an abstraction which defines a logical set of Pods and a policy by which to access them - sometimes called a micro-service. You can use services for cross-operating system connectivity. In Windows, services can utilize the following types, properties and capabilities:

    • Service Environment variables
    • NodePort
    • ClusterIP
    • LoadBalancer
    • ExternalName
    • Headless services

Pods, Controllers and Services are critical elements to managing Windows workloads on Kubernetes. However, on their own they are not enough to enable the proper lifecycle management of Windows workloads in a dynamic cloud native environment. We added support for the following features:

  • Pod and container metrics
  • Horizontal Pod Autoscaler support
  • kubectl Exec
  • Resource Quotas
  • Scheduler preemption

Container Runtime

Docker EE
FEATURE STATE: Kubernetes v1.14 [stable]

Docker EE-basic 19.03+ is the recommended container runtime for all Windows Server versions. This works with the dockershim code included in the kubelet.

FEATURE STATE: Kubernetes v1.20 [stable]

ContainerD 1.4.0+ can also be used as the container runtime for Windows Kubernetes nodes.

Learn how to install ContainerD on a Windows.

Caution: There is a known limitation when using GMSA with ContainerD to access Windows network shares which requires a kernel patch. Updates to address this limitation are currently available for Windows Server, Version 2004 and will be available for Windows Server 2019 in early 2021. Check for updates on the Microsoft Windows Containers issue tracker.

Persistent Storage

Kubernetes volumes enable complex applications, with data persistence and Pod volume sharing requirements, to be deployed on Kubernetes. Management of persistent volumes associated with a specific storage back-end or protocol includes actions such as: provisioning/de-provisioning/resizing of volumes, attaching/detaching a volume to/from a Kubernetes node and mounting/dismounting a volume to/from individual containers in a pod that needs to persist data. The code implementing these volume management actions for a specific storage back-end or protocol is shipped in the form of a Kubernetes volume plugin. The following broad classes of Kubernetes volume plugins are supported on Windows:

In-tree Volume Plugins

Code associated with in-tree volume plugins ship as part of the core Kubernetes code base. Deployment of in-tree volume plugins do not require installation of additional scripts or deployment of separate containerized plugin components. These plugins can handle: provisioning/de-provisioning and resizing of volumes in the storage backend, attaching/detaching of volumes to/from a Kubernetes node and mounting/dismounting a volume to/from individual containers in a pod. The following in-tree plugins support Windows nodes:

FlexVolume Plugins

Code associated with FlexVolume plugins ship as out-of-tree scripts or binaries that need to be deployed directly on the host. FlexVolume plugins handle attaching/detaching of volumes to/from a Kubernetes node and mounting/dismounting a volume to/from individual containers in a pod. Provisioning/De-provisioning of persistent volumes associated with FlexVolume plugins may be handled through an external provisioner that is typically separate from the FlexVolume plugins. The following FlexVolume plugins, deployed as powershell scripts on the host, support Windows nodes:

CSI Plugins
FEATURE STATE: Kubernetes v1.19 [beta]

Code associated with CSI plugins ship as out-of-tree scripts and binaries that are typically distributed as container images and deployed using standard Kubernetes constructs like DaemonSets and StatefulSets. CSI plugins handle a wide range of volume management actions in Kubernetes: provisioning/de-provisioning/resizing of volumes, attaching/detaching of volumes to/from a Kubernetes node and mounting/dismounting a volume to/from individual containers in a pod, backup/restore of persistent data using snapshots and cloning. CSI plugins typically consist of node plugins (that run on each node as a DaemonSet) and controller plugins.

CSI node plugins (especially those associated with persistent volumes exposed as either block devices or over a shared file-system) need to perform various privileged operations like scanning of disk devices, mounting of file systems, etc. These operations differ for each host operating system. For Linux worker nodes, containerized CSI node plugins are typically deployed as privileged containers. For Windows worker nodes, privileged operations for containerized CSI node plugins is supported using csi-proxy, a community-managed, stand-alone binary that needs to be pre-installed on each Windows node. Please refer to the deployment guide of the CSI plugin you wish to deploy for further details.


Networking for Windows containers is exposed through CNI plugins. Windows containers function similarly to virtual machines in regards to networking. Each container has a virtual network adapter (vNIC) which is connected to a Hyper-V virtual switch (vSwitch). The Host Networking Service (HNS) and the Host Compute Service (HCS) work together to create containers and attach container vNICs to networks. HCS is responsible for the management of containers whereas HNS is responsible for the management of networking resources such as:

  • Virtual networks (including creation of vSwitches)
  • Endpoints / vNICs
  • Namespaces
  • Policies (Packet encapsulations, Load-balancing rules, ACLs, NAT'ing rules, etc.)

The following service spec types are supported:

  • NodePort
  • ClusterIP
  • LoadBalancer
  • ExternalName
Network modes

Windows supports five different networking drivers/modes: L2bridge, L2tunnel, Overlay, Transparent, and NAT. In a heterogeneous cluster with Windows and Linux worker nodes, you need to select a networking solution that is compatible on both Windows and Linux. The following out-of-tree plugins are supported on Windows, with recommendations on when to use each CNI:

Network DriverDescriptionContainer Packet ModificationsNetwork PluginsNetwork Plugin Characteristics
L2bridgeContainers are attached to an external vSwitch. Containers are attached to the underlay network, although the physical network doesn't need to learn the container MACs because they are rewritten on ingress/egress.MAC is rewritten to host MAC, IP may be rewritten to host IP using HNS OutboundNAT policy.win-bridge, Azure-CNI, Flannel host-gateway uses win-bridgewin-bridge uses L2bridge network mode, connects containers to the underlay of hosts, offering best performance. Requires user-defined routes (UDR) for inter-node connectivity.
L2TunnelThis is a special case of l2bridge, but only used on Azure. All packets are sent to the virtualization host where SDN policy is applied.MAC rewritten, IP visible on the underlay networkAzure-CNIAzure-CNI allows integration of containers with Azure vNET, and allows them to leverage the set of capabilities that Azure Virtual Network provides. For example, securely connect to Azure services or use Azure NSGs. See azure-cni for some examples
Overlay (Overlay networking for Windows in Kubernetes is in alpha stage)Containers are given a vNIC connected to an external vSwitch. Each overlay network gets its own IP subnet, defined by a custom IP prefix.The overlay network driver uses VXLAN encapsulation.Encapsulated with an outer header.Win-overlay, Flannel VXLAN (uses win-overlay)win-overlay should be used when virtual container networks are desired to be isolated from underlay of hosts (e.g. for security reasons). Allows for IPs to be re-used for different overlay networks (which have different VNID tags) if you are restricted on IPs in your datacenter. This option requires KB4489899 on Windows Server 2019.
Transparent (special use case for ovn-kubernetes)Requires an external vSwitch. Containers are attached to an external vSwitch which enables intra-pod communication via logical networks (logical switches and routers).Packet is encapsulated either via GENEVE or STT tunneling to reach pods which are not on the same host.
Packets are forwarded or dropped via the tunnel metadata information supplied by the ovn network controller.
NAT is done for north-south communication.
ovn-kubernetesDeploy via ansible. Distributed ACLs can be applied via Kubernetes policies. IPAM support. Load-balancing can be achieved without kube-proxy. NATing is done without using iptables/netsh.
NAT (not used in Kubernetes)Containers are given a vNIC connected to an internal vSwitch. DNS/DHCP is provided using an internal component called WinNATMAC and IP is rewritten to host MAC/IP.natIncluded here for completeness

As outlined above, the Flannel CNI meta plugin is also supported on Windows via the VXLAN network backend (alpha support ; delegates to win-overlay) and host-gateway network backend (stable support; delegates to win-bridge). This plugin supports delegating to one of the reference CNI plugins (win-overlay, win-bridge), to work in conjunction with Flannel daemon on Windows (Flanneld) for automatic node subnet lease assignment and HNS network creation. This plugin reads in its own configuration file (cni.conf), and aggregates it with the environment variables from the FlannelD generated subnet.env file. It then delegates to one of the reference CNI plugins for network plumbing, and sends the correct configuration containing the node-assigned subnet to the IPAM plugin (e.g. host-local).

For the node, pod, and service objects, the following network flows are supported for TCP/UDP traffic:

  • Pod -> Pod (IP)
  • Pod -> Pod (Name)
  • Pod -> Service (Cluster IP)
  • Pod -> Service (PQDN, but only if there are no ".")
  • Pod -> Service (FQDN)
  • Pod -> External (IP)
  • Pod -> External (DNS)
  • Node -> Pod
  • Pod -> Node
IP address management (IPAM)

The following IPAM options are supported on Windows:

Load balancing and Services

On Windows, you can use the following settings to configure Services and load balancing behavior:

Windows Service Settings
FeatureDescriptionSupported Kubernetes versionSupported Windows OS buildHow to enable
Session affinityEnsures that connections from a particular client are passed to the same Pod each time.v1.19+Windows Server vNext Insider Preview Build 19551 (or higher)Set service.spec.sessionAffinity to "ClientIP"
Direct Server ReturnLoad balancing mode where the IP address fixups and the LBNAT occurs at the container vSwitch port directly; service traffic arrives with the source IP set as the originating pod IP. Promises lower latency and scalability.v1.15+Windows Server, version 2004Set the following flags in kube-proxy: --feature-gates="WinDSR=true" --enable-dsr=true
Preserve-DestinationSkips DNAT of service traffic, thereby preserving the virtual IP of the target service in packets reaching the backend Pod. This setting will also ensure that the client IP of incoming packets get preserved.v1.15+Windows Server, version 1903 (or higher)Set "preserve-destination": "true" in service annotations and enable DSR flags in kube-proxy.
IPv4/IPv6 dual-stack networkingNative IPv4-to-IPv4 in parallel with IPv6-to-IPv6 communications to, from, and within a clusterv1.19+Windows Server vNext Insider Preview Build 19603 (or higher)See IPv4/IPv6 dual-stack

IPv4/IPv6 dual-stack

You can enable IPv4/IPv6 dual-stack networking for l2bridge networks using the IPv6DualStack feature gate. See enable IPv4/IPv6 dual stack for more details.

Note: On Windows, using IPv6 with Kubernetes require Windows Server, version 2004 (kernel version 10.0.19041.610) or later.
Note: Overlay (VXLAN) networks on Windows do not support dual-stack networking today.


Windows is only supported as a worker node in the Kubernetes architecture and component matrix. This means that a Kubernetes cluster must always include Linux master nodes, zero or more Linux worker nodes, and zero or more Windows worker nodes.

Resource Handling

Linux cgroups are used as a pod boundary for resource controls in Linux. Containers are created within that boundary for network, process and file system isolation. The cgroups APIs can be used to gather cpu/io/memory stats. In contrast, Windows uses a Job object per container with a system namespace filter to contain all processes in a container and provide logical isolation from the host. There is no way to run a Windows container without the namespace filtering in place. This means that system privileges cannot be asserted in the context of the host, and thus privileged containers are not available on Windows. Containers cannot assume an identity from the host because the Security Account Manager (SAM) is separate.

Resource Reservations

Memory Reservations

Windows does not have an out-of-memory process killer as Linux does. Windows always treats all user-mode memory allocations as virtual, and pagefiles are mandatory. The net effect is that Windows won't reach out of memory conditions the same way Linux does, and processes page to disk instead of being subject to out of memory (OOM) termination. If memory is over-provisioned and all physical memory is exhausted, then paging can slow down performance.

Keeping memory usage within reasonable bounds is possible using the kubelet parameters --kubelet-reserve and/or --system-reserve to account for memory usage on the node (outside of containers). This reduces NodeAllocatable.

Note: As you deploy workloads, use resource limits (must set only limits or limits must equal requests) on containers. This also subtracts from NodeAllocatable and prevents the scheduler from adding more pods once a node is full.

A best practice to avoid over-provisioning is to configure the kubelet with a system reserved memory of at least 2GB to account for Windows, Docker, and Kubernetes processes.

CPU Reservations

To account for Windows, Docker and other Kubernetes host processes it is recommended to reserve a percentage of CPU so they are able to respond to events. This value needs to be scaled based on the number of CPU cores available on the Windows node.To determine this percentage a user should identify the maximum pod density for each of their nodes and monitor the CPU usage of the system services choosing a value that meets their workload needs.

Keeping CPU usage within reasonable bounds is possible using the kubelet parameters --kubelet-reserve and/or --system-reserve to account for CPU usage on the node (outside of containers). This reduces NodeAllocatable.

Feature Restrictions

  • TerminationGracePeriod: not implemented
  • Single file mapping: to be implemented with CRI-ContainerD
  • Termination message: to be implemented with CRI-ContainerD
  • Privileged Containers: not currently supported in Windows containers
  • HugePages: not currently supported in Windows containers
  • The existing node problem detector is Linux-only and requires privileged containers. In general, we don't expect this to be used on Windows because privileged containers are not supported
  • Not all features of shared namespaces are supported (see API section for more details)

Difference in behavior of flags when compared to Linux

The behavior of the following kubelet flags is different on Windows nodes as described below:

  • --kubelet-reserve, --system-reserve , and --eviction-hard flags update Node Allocatable
  • Eviction by using --enforce-node-allocable is not implemented
  • Eviction by using --eviction-hard and --eviction-soft are not implemented
  • MemoryPressure Condition is not implemented
  • There are no OOM eviction actions taken by the kubelet
  • Kubelet running on the windows node does not have memory restrictions. --kubelet-reserve and --system-reserve do not set limits on kubelet or processes running on the host. This means kubelet or a process on the host could cause memory resource starvation outside the node-allocatable and scheduler
  • An additional flag to set the priority of the kubelet process is available on the Windows nodes called --windows-priorityclass. This flag allows kubelet process to get more CPU time slices when compared to other processes running on the Windows host. More information on the allowable values and their meaning is available at Windows Priority Classes. In order for kubelet to always have enough CPU cycles it is recommended to set this flag to ABOVE_NORMAL_PRIORITY_CLASS and above


Windows has a layered filesystem driver to mount container layers and create a copy filesystem based on NTFS. All file paths in the container are resolved only within the context of that container.

  • With Docker Volume mounts can only target a directory in the container, and not an individual file. This limitation does not exist with CRI-containerD.
  • Volume mounts cannot project files or directories back to the host filesystem
  • Read-only filesystems are not supported because write access is always required for the Windows registry and SAM database. However, read-only volumes are supported
  • Volume user-masks and permissions are not available. Because the SAM is not shared between the host & container, there's no mapping between them. All permissions are resolved within the context of the container

As a result, the following storage functionality is not supported on Windows nodes

  • Volume subpath mounts. Only the entire volume can be mounted in a Windows container.
  • Subpath volume mounting for Secrets
  • Host mount projection
  • DefaultMode (due to UID/GID dependency)
  • Read-only root filesystem. Mapped volumes still support readOnly
  • Block device mapping
  • Memory as the storage medium
  • File system features like uui/guid, per-user Linux filesystem permissions
  • NFS based storage/volume support
  • Expanding the mounted volume (resizefs)


Windows Container Networking differs in some important ways from Linux networking. The Microsoft documentation for Windows Container Networking contains additional details and background.

The Windows host networking service and virtual switch implement namespacing and can create virtual NICs as needed for a pod or container. However, many configurations such as DNS, routes, and metrics are stored in the Windows registry database rather than /etc/... files as they are on Linux. The Windows registry for the container is separate from that of the host, so concepts like mapping /etc/resolv.conf from the host into a container don't have the same effect they would on Linux. These must be configured using Windows APIs run in the context of that container. Therefore CNI implementations need to call the HNS instead of relying on file mappings to pass network details into the pod or container.

The following networking functionality is not supported on Windows nodes

  • Host networking mode is not available for Windows pods
  • Local NodePort access from the node itself fails (works for other nodes or external clients)
  • Accessing service VIPs from nodes will be available with a future release of Windows Server
  • A single service can only support up to 64 backend pods / unique destination IPs
  • Overlay networking support in kube-proxy is an alpha release. In addition, it requires KB4482887 to be installed on Windows Server 2019
  • Local Traffic Policy and DSR mode
  • Windows containers connected to l2bridge, l2tunnel, or overlay networks do not support communicating over the IPv6 stack. There is outstanding Windows platform work required to enable these network drivers to consume IPv6 addresses and subsequent Kubernetes work in kubelet, kube-proxy, and CNI plugins.
  • Outbound communication using the ICMP protocol via the win-overlay, win-bridge, and Azure-CNI plugin. Specifically, the Windows data plane (VFP) doesn't support ICMP packet transpositions. This means:
    • ICMP packets directed to destinations within the same network (e.g. pod to pod communication via ping) work as expected and without any limitations
    • TCP/UDP packets work as expected and without any limitations
    • ICMP packets directed to pass through a remote network (e.g. pod to external internet communication via ping) cannot be transposed and thus will not be routed back to their source
    • Since TCP/UDP packets can still be transposed, one can substitute ping <destination> with curl <destination> to be able to debug connectivity to the outside world.

These features were added in Kubernetes v1.15:

  • kubectl port-forward
CNI Plugins
  • Windows reference network plugins win-bridge and win-overlay do not currently implement CNI spec v0.4.0 due to missing "CHECK" implementation.
  • The Flannel VXLAN CNI has the following limitations on Windows:
  1. Node-pod connectivity isn't possible by design. It's only possible for local pods with Flannel v0.12.0 (or higher).
  2. We are restricted to using VNI 4096 and UDP port 4789. The VNI limitation is being worked on and will be overcome in a future release (open-source flannel changes). See the official Flannel VXLAN backend docs for more details on these parameters.
  • ClusterFirstWithHostNet is not supported for DNS. Windows treats all names with a '.' as a FQDN and skips PQDN resolution
  • On Linux, you have a DNS suffix list, which is used when trying to resolve PQDNs. On Windows, we only have 1 DNS suffix, which is the DNS suffix associated with that pod's namespace (mydns.svc.cluster.local for example). Windows can resolve FQDNs and services or names resolvable with only that suffix. For example, a pod spawned in the default namespace, will have the DNS suffix default.svc.cluster.local. On a Windows pod, you can resolve both kubernetes.default.svc.cluster.local and kubernetes, but not the in-betweens, like kubernetes.default or kubernetes.default.svc.
  • On Windows, there are multiple DNS resolvers that can be used. As these come with slightly different behaviors, using the Resolve-DNSName utility for name query resolutions is recommended.

Kubernetes on Windows does not support single-stack "IPv6-only" networking. However,dual-stack IPv4/IPv6 networking for pods and nodes with single-family services is supported. See IPv4/IPv6 dual-stack networking for more details.

Session affinity

Setting the maximum session sticky time for Windows services using service.spec.sessionAffinityConfig.clientIP.timeoutSeconds is not supported.


Secrets are written in clear text on the node's volume (as compared to tmpfs/in-memory on linux). This means customers have to do two things

  1. Use file ACLs to secure the secrets file location
  2. Use volume-level encryption using BitLocker

RunAsUsername can be specified for Windows Pod's or Container's to execute the Container processes as a node-default user. This is roughly equivalent to RunAsUser.

Linux specific pod security context privileges such as SELinux, AppArmor, Seccomp, Capabilities (POSIX Capabilities), and others are not supported.

In addition, as mentioned already, privileged containers are not supported on Windows.


There are no differences in how most of the Kubernetes APIs work for Windows. The subtleties around what's different come down to differences in the OS and container runtime. In certain situations, some properties on workload APIs such as Pod or Container were designed with an assumption that they are implemented on Linux, failing to run on Windows.

At a high level, these OS concepts are different:

  • Identity - Linux uses userID (UID) and groupID (GID) which are represented as integer types. User and group names are not canonical - they are an alias in /etc/groups or /etc/passwd back to UID+GID. Windows uses a larger binary security identifier (SID) which is stored in the Windows Security Access Manager (SAM) database. This database is not shared between the host and containers, or between containers.
  • File permissions - Windows uses an access control list based on SIDs, rather than a bitmask of permissions and UID+GID
  • File paths - convention on Windows is to use \ instead of /. The Go IO libraries accept both types of file path separators. However, when you're setting a path or command line that's interpreted inside a container, \ may be needed.
  • Signals - Windows interactive apps handle termination differently, and can implement one or more of these:
    • A UI thread handles well-defined messages including WM_CLOSE
    • Console apps handle ctrl-c or ctrl-break using a Control Handler
    • Services register a Service Control Handler function that can accept SERVICE_CONTROL_STOP control codes

Exit Codes follow the same convention where 0 is success, nonzero is failure. The specific error codes may differ across Windows and Linux. However, exit codes passed from the Kubernetes components (kubelet, kube-proxy) are unchanged.

  • V1.Container.ResourceRequirements.limits.cpu and V1.Container.ResourceRequirements.limits.memory - Windows doesn't use hard limits for CPU allocations. Instead, a share system is used. The existing fields based on millicores are scaled into relative shares that are followed by the Windows scheduler. see: kuberuntime/helpers_windows.go, see: resource controls in Microsoft docs
    • Huge pages are not implemented in the Windows container runtime, and are not available. They require asserting a user privilege that's not configurable for containers.
  • V1.Container.ResourceRequirements.requests.cpu and V1.Container.ResourceRequirements.requests.memory - Requests are subtracted from node available resources, so they can be used to avoid overprovisioning a node. However, they cannot be used to guarantee resources in an overprovisioned node. They should be applied to all containers as a best practice if the operator wants to avoid overprovisioning entirely.
  • V1.Container.SecurityContext.allowPrivilegeEscalation - not possible on Windows, none of the capabilities are hooked up
  • V1.Container.SecurityContext.Capabilities - POSIX capabilities are not implemented on Windows
  • V1.Container.SecurityContext.privileged - Windows doesn't support privileged containers
  • V1.Container.SecurityContext.procMount - Windows doesn't have a /proc filesystem
  • V1.Container.SecurityContext.readOnlyRootFilesystem - not possible on Windows, write access is required for registry & system processes to run inside the container
  • V1.Container.SecurityContext.runAsGroup - not possible on Windows, no GID support
  • V1.Container.SecurityContext.runAsNonRoot - Windows does not have a root user. The closest equivalent is ContainerAdministrator which is an identity that doesn't exist on the node.
  • V1.Container.SecurityContext.runAsUser - not possible on Windows, no UID support as int.
  • V1.Container.SecurityContext.seLinuxOptions - not possible on Windows, no SELinux
  • V1.Container.terminationMessagePath - this has some limitations in that Windows doesn't support mapping single files. The default value is /dev/termination-log, which does work because it does not exist on Windows by default.
  • V1.Pod.hostIPC, v1.pod.hostpid - host namespace sharing is not possible on Windows
  • V1.Pod.hostNetwork - There is no Windows OS support to share the host network
  • V1.Pod.dnsPolicy - ClusterFirstWithHostNet - is not supported because Host Networking is not supported on Windows.
  • V1.Pod.podSecurityContext - see V1.PodSecurityContext below
  • V1.Pod.shareProcessNamespace - this is a beta feature, and depends on Linux namespaces which are not implemented on Windows. Windows cannot share process namespaces or the container's root filesystem. Only the network can be shared.
  • V1.Pod.terminationGracePeriodSeconds - this is not fully implemented in Docker on Windows, see: reference. The behavior today is that the ENTRYPOINT process is sent CTRL_SHUTDOWN_EVENT, then Windows waits 5 seconds by default, and finally shuts down all processes using the normal Windows shutdown behavior. The 5 second default is actually in the Windows registry inside the container, so it can be overridden when the container is built.
  • V1.Pod.volumeDevices - this is a beta feature, and is not implemented on Windows. Windows cannot attach raw block devices to pods.
  • V1.Pod.volumes - EmptyDir, Secret, ConfigMap, HostPath - all work and have tests in TestGrid
    • V1.emptyDirVolumeSource - the Node default medium is disk on Windows. Memory is not supported, as Windows does not have a built-in RAM disk.
  • V1.VolumeMount.mountPropagation - mount propagation is not supported on Windows.

None of the PodSecurityContext fields work on Windows. They're listed here for reference.

  • V1.PodSecurityContext.SELinuxOptions - SELinux is not available on Windows
  • V1.PodSecurityContext.RunAsUser - provides a UID, not available on Windows
  • V1.PodSecurityContext.RunAsGroup - provides a GID, not available on Windows
  • V1.PodSecurityContext.RunAsNonRoot - Windows does not have a root user. The closest equivalent is ContainerAdministrator which is an identity that doesn't exist on the node.
  • V1.PodSecurityContext.SupplementalGroups - provides GID, not available on Windows
  • V1.PodSecurityContext.Sysctls - these are part of the Linux sysctl interface. There's no equivalent on Windows.

Operating System Version Restrictions

Windows has strict compatibility rules, where the host OS version must match the container base image OS version. Only Windows containers with a container operating system of Windows Server 2019 are supported. Hyper-V isolation of containers, enabling some backward compatibility of Windows container image versions, is planned for a future release.

Getting Help and Troubleshooting

Your main source of help for troubleshooting your Kubernetes cluster should start with this section. Some additional, Windows-specific troubleshooting help is included in this section. Logs are an important element of troubleshooting issues in Kubernetes. Make sure to include them any time you seek troubleshooting assistance from other contributors. Follow the instructions in the SIG-Windows contributing guide on gathering logs.

  1. How do I know start.ps1 completed successfully?

    You should see kubelet, kube-proxy, and (if you chose Flannel as your networking solution) flanneld host-agent processes running on your node, with running logs being displayed in separate PowerShell windows. In addition to this, your Windows node should be listed as "Ready" in your Kubernetes cluster.

  2. Can I configure the Kubernetes node processes to run in the background as services?

    Kubelet and kube-proxy are already configured to run as native Windows Services, offering resiliency by re-starting the services automatically in the event of failure (for example a process crash). You have two options for configuring these node components as services.

    1. As native Windows Services

      Kubelet & kube-proxy can be run as native Windows Services using sc.exe.

      # Create the services for kubelet and kube-proxy in two separate commands
      sc.exe create <component_name> binPath= "<path_to_binary> --service <other_args>"
      # Please note that if the arguments contain spaces, they must be escaped.
      sc.exe create kubelet binPath= "C:\kubelet.exe --service --hostname-override 'minion' <other_args>"
      # Start the services
      Start-Service kubelet
      Start-Service kube-proxy
      # Stop the service
      Stop-Service kubelet (-Force)
      Stop-Service kube-proxy (-Force)
      # Query the service status
      Get-Service kubelet
      Get-Service kube-proxy
    2. Using nssm.exe

      You can also always use alternative service managers like nssm.exe to run these processes (flanneld, kubelet & kube-proxy) in the background for you. You can use this sample script, leveraging nssm.exe to register kubelet, kube-proxy, and flanneld.exe to run as Windows services in the background.

      register-svc.ps1 -NetworkMode <Network mode> -ManagementIP <Windows Node IP> -ClusterCIDR <Cluster subnet> -KubeDnsServiceIP <Kube-dns Service IP> -LogDir <Directory to place logs>
      # NetworkMode      = The network mode l2bridge (flannel host-gw, also the default value) or overlay (flannel vxlan) chosen as a network solution
      # ManagementIP     = The IP address assigned to the Windows node. You can use ipconfig to find this
      # ClusterCIDR      = The cluster subnet range. (Default value
      # KubeDnsServiceIP = The Kubernetes DNS service IP (Default value
      # LogDir           = The directory where kubelet and kube-proxy logs are redirected into their respective output files (Default value C:\k)

      If the above referenced script is not suitable, you can manually configure nssm.exe using the following examples.

      # Register flanneld.exe
      nssm install flanneld C:\flannel\flanneld.exe
      nssm set flanneld AppParameters --kubeconfig-file=c:\k\config --iface=<ManagementIP> --ip-masq=1 --kube-subnet-mgr=1
      nssm set flanneld AppEnvironmentExtra NODE_NAME=<hostname>
      nssm set flanneld AppDirectory C:\flannel
      nssm start flanneld
      # Register kubelet.exe
      # Microsoft releases the pause infrastructure container at mcr.microsoft.com/oss/kubernetes/pause:1.4.1
      nssm install kubelet C:\k\kubelet.exe
      nssm set kubelet AppParameters --hostname-override=<hostname> --v=6 --pod-infra-container-image=mcr.microsoft.com/oss/kubernetes/pause:1.4.1 --resolv-conf="" --allow-privileged=true --enable-debugging-handlers --cluster-dns=<DNS-service-IP> --cluster-domain=cluster.local --kubeconfig=c:\k\config --hairpin-mode=promiscuous-bridge --image-pull-progress-deadline=20m --cgroups-per-qos=false  --log-dir=<log directory> --logtostderr=false --enforce-node-allocatable="" --network-plugin=cni --cni-bin-dir=c:\k\cni --cni-conf-dir=c:\k\cni\config
      nssm set kubelet AppDirectory C:\k
      nssm start kubelet
      # Register kube-proxy.exe (l2bridge / host-gw)
      nssm install kube-proxy C:\k\kube-proxy.exe
      nssm set kube-proxy AppDirectory c:\k
      nssm set kube-proxy AppParameters --v=4 --proxy-mode=kernelspace --hostname-override=<hostname>--kubeconfig=c:\k\config --enable-dsr=false --log-dir=<log directory> --logtostderr=false
      nssm.exe set kube-proxy AppEnvironmentExtra KUBE_NETWORK=cbr0
      nssm set kube-proxy DependOnService kubelet
      nssm start kube-proxy
      # Register kube-proxy.exe (overlay / vxlan)
      nssm install kube-proxy C:\k\kube-proxy.exe
      nssm set kube-proxy AppDirectory c:\k
      nssm set kube-proxy AppParameters --v=4 --proxy-mode=kernelspace --feature-gates="WinOverlay=true" --hostname-override=<hostname> --kubeconfig=c:\k\config --network-name=vxlan0 --source-vip=<source-vip> --enable-dsr=false --log-dir=<log directory> --logtostderr=false
      nssm set kube-proxy DependOnService kubelet
      nssm start kube-proxy

      For initial troubleshooting, you can use the following flags in nssm.exe to redirect stdout and stderr to a output file:

      nssm set <Service Name> AppStdout C:\k\mysvc.log
      nssm set <Service Name> AppStderr C:\k\mysvc.log

      For additional details, see official nssm usage docs.

  3. My Windows Pods do not have network connectivity

    If you are using virtual machines, ensure that MAC spoofing is enabled on all the VM network adapter(s).

  4. My Windows Pods cannot ping external resources

    Windows Pods do not have outbound rules programmed for the ICMP protocol today. However, TCP/UDP is supported. When trying to demonstrate connectivity to resources outside of the cluster, please substitute ping <IP> with corresponding curl <IP> commands.

    If you are still facing problems, most likely your network configuration in cni.conf deserves some extra attention. You can always edit this static file. The configuration update will apply to any newly created Kubernetes resources.

    One of the Kubernetes networking requirements (see Kubernetes model) is for cluster communication to occur without NAT internally. To honor this requirement, there is an ExceptionList for all the communication where we do not want outbound NAT to occur. However, this also means that you need to exclude the external IP you are trying to query from the ExceptionList. Only then will the traffic originating from your Windows pods be SNAT'ed correctly to receive a response from the outside world. In this regard, your ExceptionList in cni.conf should look as follows:

    "ExceptionList": [
                    "",  # Cluster subnet
                    "",   # Service subnet
                    "" # Management (host) subnet
  5. My Windows node cannot access NodePort service

    Local NodePort access from the node itself fails. This is a known limitation. NodePort access works from other nodes or external clients.

  6. vNICs and HNS endpoints of containers are being deleted

    This issue can be caused when the hostname-override parameter is not passed to kube-proxy. To resolve it, users need to pass the hostname to kube-proxy as follows:

    C:\k\kube-proxy.exe --hostname-override=$(hostname)
  7. With flannel my nodes are having issues after rejoining a cluster

    Whenever a previously deleted node is being re-joined to the cluster, flannelD tries to assign a new pod subnet to the node. Users should remove the old pod subnet configuration files in the following paths:

    Remove-Item C:\k\SourceVip.json
    Remove-Item C:\k\SourceVipRequest.json
  8. After launching start.ps1, flanneld is stuck in "Waiting for the Network to be created"

    There are numerous reports of this issue; most likely it is a timing issue for when the management IP of the flannel network is set. A workaround is to relaunch start.ps1 or relaunch it manually as follows:

    PS C:> [Environment]::SetEnvironmentVariable("NODE_NAME", "<Windows_Worker_Hostname>")
    PS C:> C:\flannel\flanneld.exe --kubeconfig-file=c:\k\config --iface=<Windows_Worker_Node_IP> --ip-masq=1 --kube-subnet-mgr=1
  9. My Windows Pods cannot launch because of missing /run/flannel/subnet.env

    This indicates that Flannel didn't launch correctly. You can either try to restart flanneld.exe or you can copy the files over manually from /run/flannel/subnet.env on the Kubernetes master to C:\run\flannel\subnet.env on the Windows worker node and modify the FLANNEL_SUBNET row to a different number. For example, if node subnet is desired:

  10. My Windows node cannot access my services using the service IP

    This is a known limitation of the current networking stack on Windows. Windows Pods are able to access the service IP however.

  11. No network adapter is found when starting kubelet

    The Windows networking stack needs a virtual adapter for Kubernetes networking to work. If the following commands return no results (in an admin shell), virtual network creation — a necessary prerequisite for Kubelet to work — has failed:

    Get-HnsNetwork | ? Name -ieq "cbr0"
    Get-NetAdapter | ? Name -Like "vEthernet (Ethernet*"

    Often it is worthwhile to modify the InterfaceName parameter of the start.ps1 script, in cases where the host's network adapter isn't "Ethernet". Otherwise, consult the output of the start-kubelet.ps1 script to see if there are errors during virtual network creation.

  12. My Pods are stuck at "Container Creating" or restarting over and over

    Check that your pause image is compatible with your OS version. The instructions assume that both the OS and the containers are version 1803. If you have a later version of Windows, such as an Insider build, you need to adjust the images accordingly. Please refer to the Microsoft's Docker repository for images. Regardless, both the pause image Dockerfile and the sample service expect the image to be tagged as :latest.

  13. DNS resolution is not properly working

    Check the DNS limitations for Windows in this section.

  14. kubectl port-forward fails with "unable to do port forwarding: wincat not found"

    This was implemented in Kubernetes 1.15 by including wincat.exe in the pause infrastructure container mcr.microsoft.com/oss/kubernetes/pause:1.4.1. Be sure to use these versions or newer ones. If you would like to build your own pause infrastructure container be sure to include wincat.

  15. My Kubernetes installation is failing because my Windows Server node is behind a proxy

    If you are behind a proxy, the following PowerShell environment variables must be defined:

    [Environment]::SetEnvironmentVariable("HTTP_PROXY", "http://proxy.example.com:80/", [EnvironmentVariableTarget]::Machine)
    [Environment]::SetEnvironmentVariable("HTTPS_PROXY", "http://proxy.example.com:443/", [EnvironmentVariableTarget]::Machine)
  16. What is a pause container?

    In a Kubernetes Pod, an infrastructure or "pause" container is first created to host the container endpoint. Containers that belong to the same pod, including infrastructure and worker containers, share a common network namespace and endpoint (same IP and port space). Pause containers are needed to accommodate worker containers crashing or restarting without losing any of the networking configuration.

    The "pause" (infrastructure) image is hosted on Microsoft Container Registry (MCR). You can access it using mcr.microsoft.com/oss/kubernetes/pause:1.4.1. For more details, see the DOCKERFILE.

Further investigation

If these steps don't resolve your problem, you can get help running Windows containers on Windows nodes in Kubernetes through:

Reporting Issues and Feature Requests

If you have what looks like a bug, or you would like to make a feature request, please use the GitHub issue tracking system. You can open issues on GitHub and assign them to SIG-Windows. You should first search the list of issues in case it was reported previously and comment with your experience on the issue and add additional logs. SIG-Windows Slack is also a great avenue to get some initial support and troubleshooting ideas prior to creating a ticket.

If filing a bug, please include detailed information about how to reproduce the problem, such as:

  • Kubernetes version: kubectl version
  • Environment details: Cloud provider, OS distro, networking choice and configuration, and Docker version
  • Detailed steps to reproduce the problem
  • Relevant logs
  • Tag the issue sig/windows by commenting on the issue with /sig windows to bring it to a SIG-Windows member's attention

What's next

We have a lot of features in our roadmap. An abbreviated high level list is included below, but we encourage you to view our roadmap project and help us make Windows support better by contributing.

Hyper-V isolation

Hyper-V isolation is required to enable the following use cases for Windows containers in Kubernetes:

  • Hypervisor-based isolation between pods for additional security
  • Backwards compatibility allowing a node to run a newer Windows Server version without requiring containers to be rebuilt
  • Specific CPU/NUMA settings for a pod
  • Memory isolation and reservations

Hyper-V isolation support will be added in a later release and will require CRI-Containerd.

Deployment with kubeadm and cluster API

Kubeadm is becoming the de facto standard for users to deploy a Kubernetes cluster. Windows node support in kubeadm is currently a work-in-progress but a guide is available here. We are also making investments in cluster API to ensure Windows nodes are properly provisioned.

4.2 - Guide for scheduling Windows containers in Kubernetes

Windows applications constitute a large portion of the services and applications that run in many organizations. This guide walks you through the steps to configure and deploy a Windows container in Kubernetes.


  • Configure an example deployment to run Windows containers on the Windows node
  • (Optional) Configure an Active Directory Identity for your Pod using Group Managed Service Accounts (GMSA)

Before you begin

  • Create a Kubernetes cluster that includes a master and a worker node running Windows Server
  • It is important to note that creating and deploying services and workloads on Kubernetes behaves in much the same way for Linux and Windows containers. Kubectl commands to interface with the cluster are identical. The example in the section below is provided to jumpstart your experience with Windows containers.

Getting Started: Deploying a Windows container

To deploy a Windows container on Kubernetes, you must first create an example application. The example YAML file below creates a simple webserver application. Create a service spec named win-webserver.yaml with the contents below:

apiVersion: v1
kind: Service
  name: win-webserver
    app: win-webserver
    # the port that this service should serve on
    - port: 80
      targetPort: 80
    app: win-webserver
  type: NodePort
apiVersion: apps/v1
kind: Deployment
    app: win-webserver
  name: win-webserver
  replicas: 2
      app: win-webserver
        app: win-webserver
      name: win-webserver
      - name: windowswebserver
        image: mcr.microsoft.com/windows/servercore:ltsc2019
        - powershell.exe
        - -command
        - "<#code used from https://gist.github.com/19WAS85/5424431#> ; $$listener = New-Object System.Net.HttpListener ; $$listener.Prefixes.Add('http://*:80/') ; $$listener.Start() ; $$callerCounts = @{} ; Write-Host('Listening at http://*:80/') ; while ($$listener.IsListening) { ;$$context = $$listener.GetContext() ;$$requestUrl = $$context.Request.Url ;$$clientIP = $$context.Request.RemoteEndPoint.Address ;$$response = $$context.Response ;Write-Host '' ;Write-Host('> {0}' -f $$requestUrl) ;  ;$$count = 1 ;$$k=$$callerCounts.Get_Item($$clientIP) ;if ($$k -ne $$null) { $$count += $$k } ;$$callerCounts.Set_Item($$clientIP, $$count) ;$$ip=(Get-NetAdapter | Get-NetIpAddress); $$header='<html><body><H1>Windows Container Web Server</H1>' ;$$callerCountsString='' ;$$callerCounts.Keys | % { $$callerCountsString+='<p>IP {0} callerCount {1} ' -f $$ip[1].IPAddress,$$callerCounts.Item($$_) } ;$$footer='</body></html>' ;$$content='{0}{1}{2}' -f $$header,$$callerCountsString,$$footer ;Write-Output $$content ;$$buffer = [System.Text.Encoding]::UTF8.GetBytes($$content) ;$$response.ContentLength64 = $$buffer.Length ;$$response.OutputStream.Write($$buffer, 0, $$buffer.Length) ;$$response.Close() ;$$responseStatus = $$response.StatusCode ;Write-Host('< {0}' -f $$responseStatus)  } ; "
      kubernetes.io/os: windows
Note: Port mapping is also supported, but for simplicity in this example the container port 80 is exposed directly to the service.
  1. Check that all nodes are healthy:

    kubectl get nodes
  2. Deploy the service and watch for pod updates:

    kubectl apply -f win-webserver.yaml
    kubectl get pods -o wide -w

    When the service is deployed correctly both Pods are marked as Ready. To exit the watch command, press Ctrl+C.

  3. Check that the deployment succeeded. To verify:

    • Two containers per pod on the Windows node, use docker ps
    • Two pods listed from the Linux master, use kubectl get pods
    • Node-to-pod communication across the network, curl port 80 of your pod IPs from the Linux master to check for a web server response
    • Pod-to-pod communication, ping between pods (and across hosts, if you have more than one Windows node) using docker exec or kubectl exec
    • Service-to-pod communication, curl the virtual service IP (seen under kubectl get services) from the Linux master and from individual pods
    • Service discovery, curl the service name with the Kubernetes default DNS suffix
    • Inbound connectivity, curl the NodePort from the Linux master or machines outside of the cluster
    • Outbound connectivity, curl external IPs from inside the pod using kubectl exec
Note: Windows container hosts are not able to access the IP of services scheduled on them due to current platform limitations of the Windows networking stack. Only Windows pods are able to access service IPs.


Capturing logs from workloads

Logs are an important element of observability; they enable users to gain insights into the operational aspect of workloads and are a key ingredient to troubleshooting issues. Because Windows containers and workloads inside Windows containers behave differently from Linux containers, users had a hard time collecting logs, limiting operational visibility. Windows workloads for example are usually configured to log to ETW (Event Tracing for Windows) or push entries to the application event log. LogMonitor, an open source tool by Microsoft, is the recommended way to monitor configured log sources inside a Windows container. LogMonitor supports monitoring event logs, ETW providers, and custom application logs, piping them to STDOUT for consumption by kubectl logs <pod>.

Follow the instructions in the LogMonitor GitHub page to copy its binaries and configuration files to all your containers and add the necessary entrypoints for LogMonitor to push your logs to STDOUT.

Using configurable Container usernames

Starting with Kubernetes v1.16, Windows containers can be configured to run their entrypoints and processes with different usernames than the image defaults. The way this is achieved is a bit different from the way it is done for Linux containers. Learn more about it here.

Managing Workload Identity with Group Managed Service Accounts

Starting with Kubernetes v1.14, Windows container workloads can be configured to use Group Managed Service Accounts (GMSA). Group Managed Service Accounts are a specific type of Active Directory account that provides automatic password management, simplified service principal name (SPN) management, and the ability to delegate the management to other administrators across multiple servers. Containers configured with a GMSA can access external Active Directory Domain resources while carrying the identity configured with the GMSA. Learn more about configuring and using GMSA for Windows containers here.

Taints and Tolerations

Users today need to use some combination of taints and node selectors in order to keep Linux and Windows workloads on their respective OS-specific nodes. This likely imposes a burden only on Windows users. The recommended approach is outlined below, with one of its main goals being that this approach should not break compatibility for existing Linux workloads.

Ensuring OS-specific workloads land on the appropriate container host

Users can ensure Windows containers can be scheduled on the appropriate host using Taints and Tolerations. All Kubernetes nodes today have the following default labels:

  • kubernetes.io/os = [windows|linux]
  • kubernetes.io/arch = [amd64|arm64|...]

If a Pod specification does not specify a nodeSelector like "kubernetes.io/os": windows, it is possible the Pod can be scheduled on any host, Windows or Linux. This can be problematic since a Windows container can only run on Windows and a Linux container can only run on Linux. The best practice is to use a nodeSelector.

However, we understand that in many cases users have a pre-existing large number of deployments for Linux containers, as well as an ecosystem of off-the-shelf configurations, such as community Helm charts, and programmatic Pod generation cases, such as with Operators. In those situations, you may be hesitant to make the configuration change to add nodeSelectors. The alternative is to use Taints. Because the kubelet can set Taints during registration, it could easily be modified to automatically add a taint when running on Windows only.

For example: --register-with-taints='os=windows:NoSchedule'

By adding a taint to all Windows nodes, nothing will be scheduled on them (that includes existing Linux Pods). In order for a Windows Pod to be scheduled on a Windows node, it would need both the nodeSelector to choose Windows, and the appropriate matching toleration.

    kubernetes.io/os: windows
    node.kubernetes.io/windows-build: '10.0.17763'
    - key: "os"
      operator: "Equal"
      value: "windows"
      effect: "NoSchedule"

Handling multiple Windows versions in the same cluster

The Windows Server version used by each pod must match that of the node. If you want to use multiple Windows Server versions in the same cluster, then you should set additional node labels and nodeSelectors.

Kubernetes 1.17 automatically adds a new label node.kubernetes.io/windows-build to simplify this. If you're running an older version, then it's recommended to add this label manually to Windows nodes.

This label reflects the Windows major, minor, and build number that need to match for compatibility. Here are values used today for each Windows Server version.

Product NameBuild Number(s)
Windows Server 201910.0.17763
Windows Server version 180910.0.17763
Windows Server version 190310.0.18362

Simplifying with RuntimeClass

RuntimeClass can be used to simplify the process of using taints and tolerations. A cluster administrator can create a RuntimeClass object which is used to encapsulate these taints and tolerations.

  1. Save this file to runtimeClasses.yml. It includes the appropriate nodeSelector for the Windows OS, architecture, and version.
apiVersion: node.k8s.io/v1
kind: RuntimeClass
  name: windows-2019
handler: 'docker'
    kubernetes.io/os: 'windows'
    kubernetes.io/arch: 'amd64'
    node.kubernetes.io/windows-build: '10.0.17763'
  - effect: NoSchedule
    key: os
    operator: Equal
    value: "windows"
  1. Run kubectl create -f runtimeClasses.yml using as a cluster administrator
  2. Add runtimeClassName: windows-2019 as appropriate to Pod specs

For example:

apiVersion: apps/v1
kind: Deployment
  name: iis-2019
    app: iis-2019
  replicas: 1
      name: iis-2019
        app: iis-2019
      runtimeClassName: windows-2019
      - name: iis
        image: mcr.microsoft.com/windows/servercore/iis:windowsservercore-ltsc2019
            cpu: 1
            memory: 800Mi
            cpu: .1
            memory: 300Mi
          - containerPort: 80
      app: iis-2019
apiVersion: v1
kind: Service
  name: iis
  type: LoadBalancer
  - protocol: TCP
    port: 80
    app: iis-2019