This document describes how to run Kubernetes using rkt as the container runtime.
Note: This document describes how to use what is known as “rktnetes”. In future, Kubernetes will support the rkt runtime through the Container Runtime Interface (CRI). At present the rkt shim for the CRI is considered “experimental”, but if you wish to use it you will find instructions in the kubeadm reference.
Systemd must be installed and enabled. The minimum systemd version required for Kubernetes v1.3 is 219
. Systemd is used to monitor and manage the pods on each node.
Install the latest rkt release. The minimum rkt version required is v1.13.0. The CoreOS Linux alpha channel ships with a recent rkt release, and you can easily upgrade rkt on CoreOS, if necessary.
The rkt API service must be running on the node.
You will need kubelet installed on the node, and it’s recommended that you run kube-proxy on all nodes. This document describes how to set the parameters for kubelet so that it uses rkt as the runtime.
You can configure Kubernetes pod networking with the usual Container Network Interface (CNI) network plugins by setting the kubelet’s --network-plugin
and --network-plugin-dir
options appropriately. Configured in this fashion, the rkt container engine will be unaware of network details, and expects to connect pods to the provided subnet.
The kubenet
plugin can be selected with the kubelet option --network-plugin=kubenet
. This plugin is currently only supported on GCE. When using kubenet, Kubernetes CNI creates and manages the network, and rkt is provided with a subnet from a bridge device connected to the GCE network.
Rather than delegating pod networking to Kubernetes, rkt can configure connectivity directly with its own contained network on a subnet provided by a bridge device, the flannel SDN, or another CNI plugin. Configured this way, rkt looks in its config directories, usually /etc/rkt/net.d
, to discover the CNI configuration and invoke the appropriate plugins to create the pod network.
The contained network is rkt’s default, so you can leave the kubelet’s --network-plugin
option empty to select this network. The contained network can be backed by any CNI plugin. With the contained network, rkt will attempt to join pods to a network named rkt.kubernetes.io
, so this network name must be used for whatever desired CNI configuration.
When using the contained network, create a network configuration file beneath the rkt network config directory that defines how to create this rkt.kubernetes.io
network in your environment. This example sets up a bridge device with the bridge
CNI plugin:
$ cat <<EOF >/etc/rkt/net.d/k8s_network_example.conf
{
"name": "rkt.kubernetes.io",
"type": "bridge",
"bridge": "mybridge",
"mtu": 1460,
"addIf": "true",
"isGateway": true,
"ipMasq": true,
"ipam": {
"type": "host-local",
"subnet": "10.22.0.0/16",
"gateway": "10.22.0.1",
"routes": [
{ "dst": "0.0.0.0/0" }
]
}
}
EOF
While it is recommended to operate flannel through the Kubernetes CNI support, you can alternatively configure the flannel plugin directly to provide the subnet for rkt’s contained network. An example CNI/flannel config file looks like this:
$ cat <<EOF >/etc/rkt/net.d/k8s_flannel_example.conf
{
"name": "rkt.kubernetes.io",
"type": "flannel",
"delegate": {
"isDefaultGateway": true
}
}
EOF
For more information on flannel configuration, see the CNI/flannel README.
rkt.kubernetes.io
./etc/hosts
file will not contain the pod’s own hostname, although /etc/hostname
is populated.To use rkt as the container runtime in a local Kubernetes cluster, supply the following flags to the kubelet:
--container-runtime=rkt
Set the node’s container runtime to rkt.--rkt-api-endpoint=HOST:PORT
Set the endpoint of the rkt API service. Default: localhost:15441
.--rkt-path=PATH_TO_RKT_BINARY
Set the path of the rkt binary. Optional. If empty, look for rkt
in $PATH
.--rkt-stage1-image=STAGE1
Set the name of the stage1 image, e.g. coreos.com/rkt/stage1-coreos
. Optional. If not set, the default Linux kernel software isolation stage1 is used.If you are using the hack/local-up-cluster.sh script to launch the cluster, you can edit the environment variables CONTAINER_RUNTIME
, RKT_PATH
, and RKT_STAGE1_IMAGE
to set these flags. RKT_PATH
and RKT_STAGE1_IMAGE
are optional if rkt
is in your $PATH` with appropriate configuration.
$ export CONTAINER_RUNTIME=rkt
$ export RKT_PATH=<rkt_binary_path>
$ export RKT_STAGE1_IMAGE=<stage1-name>
Now you can launch the cluster using the local-up-cluster.sh
script:
$ hack/local-up-cluster.sh
We are also working on getting rkt working as the container runtime in minikube.
This section outlines using the kube-up
script to launch a CoreOS/rkt cluster on GCE.
Specify the OS distribution, the GCE distributor’s master project, and the instance images for the Kubernetes master and nodes. Set the KUBE_CONTAINER_RUNTIME
to rkt
:
$ export KUBE_OS_DISTRIBUTION=coreos
$ export KUBE_GCE_MASTER_PROJECT=coreos-cloud
$ export KUBE_GCE_MASTER_IMAGE=<image_id>
$ export KUBE_GCE_NODE_PROJECT=coreos-cloud
$ export KUBE_GCE_NODE_IMAGE=<image_id>
$ export KUBE_CONTAINER_RUNTIME=rkt
Optionally, set the version of rkt by setting KUBE_RKT_VERSION
:
$ export KUBE_RKT_VERSION=1.13.0
Optionally, select an alternative stage1 isolator for the container runtime by setting KUBE_RKT_STAGE1_IMAGE
:
$ export KUBE_RKT_STAGE1_IMAGE=<stage1-name>
Then you can launch the cluster with:
$ cluster/kube-up.sh
The kube-up
script is not yet supported on AWS. Instead, we recommend following the Kubernetes on AWS guide to launch a CoreOS Kubernetes cluster on AWS, then setting kubelet options as above.
After creating the cluster, you can start deploying applications. For an introductory example, deploy a simple nginx web server. Note that this example did not have to be modified for use with a “rktnetes” cluster. More examples can be found in the Kubernetes examples directory.
rkt executes containers in an interchangeable isolation environment. This facility is called the stage1 image. There are currently three supported rkt stage1 images:
systemd-nspawn
stage1, the default. Isolates running containers with Linux kernel namespaces and cgroups in a manner similar to the default container runtime.KVM
stage1, runs containers inside a KVM hypervisor-managed virtual machine. Experimental in the Kubernetes v1.3 release.fly stage1
, which isolates containers with only a chroot
, giving host-level access to mount and network namespaces for specially-privileged utilities.In addition to the three provided stage1 images, you can create your own for specific isolation requirements. If no configuration is set, the default stage1 is used. There are two ways to select a different stage1; either per-node, or per-pod:
--rkt-stage1-image
flag, which tells the kubelet the stage1 image to use for every pod on the node. For example, --rkt-stage1-image=coreos/rkt/stage1-coreos
selects the default systemd-nspawn stage1.rkt.alpha.kubernetes.io/stage1-name-override
to override the stage1 used to execute a given pod. This allows for mixing different container isolation mechanisms on the same cluster or on the same node. For example, the following (shortened) pod manifest will run its pod with the fly stage1
to give the application – the kubelet
in this case – access to the host’s namespace:apiVersion: v1
kind: Pod
metadata:
name: kubelet
namespace: kube-system
labels:
k8s-app: kubelet
annotations:
rkt.alpha.kubernetes.io/stage1-name-override: coreos.com/rkt/stage1-fly
spec:
containers:
- name: kubelet
image: quay.io/coreos/hyperkube:v1.3.0-beta.2_coreos.0
command:
- kubelet
- --api-servers=127.0.0.1:8080
- --config=/etc/kubernetes/manifests
- --allow-privileged
- --kubeconfig=/etc/kubernetes/kubeconfig
securityContext:
privileged: true
[...]
Setting the stage1 annotation could potentially give the pod root privileges. Because of this, the privileged
boolean in the pod’s securityContext
must be set to true
.
Use rkt’s contained network with the KVM stage1, because the CNI plugin driver does not yet fully support the hypervisor-based runtime.
rkt and the default node container engine have very different designs, as do rkt’s native ACI and the Docker container image format. Users may experience different behaviors when switching from one container engine to the other. More information can be found in the Kubernetes rkt notes.
Here are a few tips for troubleshooting Kubernetes with the rkt container engine:
To check the status of running pods, use the rkt subcommands rkt list
, rkt status
, and rkt image list
. See the rkt commands documentation for more information about rkt subcommands.
Check a pod’s log using journalctl
on the node. Pods are managed and named as systemd units. The pod’s unit name is formed by concatenating a k8s_
prefix with the pod UUID, in a format like k8s_${RKT_UUID}
. Find the pod’s UUID with rkt list
to assemble its service name, then ask journalctl for the logs:
$ sudo journalctl -u k8s_ad623346
By default, the log verbosity level is 2. In order to see more log messages related to rkt, set this level to 4 or above. For a local cluster, set the environment variable: LOG_LEVEL=4
.
Kubernetes provides various tools for troubleshooting and examination. More information can be found in the app troubleshooting guide.
Create an Issue Edit this Page