If you use ctr (or the containerd API) to load images into the k8s.io namespace, those images are not visible to the kubelet. This is due to an extra layer of metadata/caching taking place in the CRI layer.

This extra layer adds complexity and inconsistency into the stack, which could lead to problems in production environments. Tools developed against containerd will get an inconsistent view of the images compared to the kubelet. Operators may take action with incomplete/inconsistent information, and cause failures in their environment.

We should try to find a way to have a single source of truth for the images loaded within the namespace.