The following features either are not supported or have large caveats when using the rkt container runtime. Increasing support for these items and others, including reasonable feature parity with the default container engine, is planned through future releases.
When mounting a host volume path that does not exist, rkt will error out. Under the Docker runtime, an empty directory will be created at the referenced path.
An example of a pod which will error out:
apiVersion: v1 kind: Pod metadata: labels: name: mount-dne name: mount-dne spec: volumes: - name: does-not-exist hostPath: path: /does/not/exist containers: - name: exit image: busybox command: ["sh", "-c", "ls /test; sleep 60"] volumeMounts: - mountPath: /test name: does-not-exist
Also note that if
subPath is specified in the container’s volumeMounts and the
subPath doesn’t exist in the corresponding volume, the pod execution will fail as well.
kubectl attach command does not work under the rkt container runtime.
Because of this, some flags in
kubectl run are not supported, including:
kubectl port-forward is not supported for pods that are executed with
Currently rkt supports only per-pod volume relabeling. After relabeling, the mounted volume is shared by all Containers in the pod. There is not yet a way to make the relabeled volume accessible to only one, or some subset, of Containers in the pod. Kubernetes issue # 28187 has the details.
kubectl get logs currently cannot get logs from applications that write them to directly to
/dev/stdout. Currently such log messages are printed on the node’s console.
Init Container feature is currently not supported.
Exponential restart back-off for a failing container is currently not supported.
--experimental-nvidia-gpus flag, and related GPU features are not supported.
Under rkt, QoS classes do not adjust the
OOM Score of Containers as occurs under Docker.
Setting the hostPID or hostIPC flags on a pod is not supported.
For example, the following pod will not run correctly:
apiVersion: v1 kind: Pod metadata: labels: name: host-ipc-pid name: host-ipc-pid spec: hostIPC: true hostPID: true containers: ...
On the other hand, when running the pod with stage1-fly, the pod will be run in the host namespace.
Patching a pod to change the image will result in the entire pod restarting, not just the container that was changed.
When the container’s image pull policy is
Always, rkt will always pull the image from remote even if the image has not changed at all.
This can add significant latency for large images.
The issue is tracked by rkt upstream at #2937.