Containerized services have no resource limits after they restart

[dsseng]

Bug Report

Description

If a service like Kubelet restarts, containerd (system engine) removes its cgroup and then re-creates it, without keeping limits configured by machined.

A similar problem exists for services started by the Process runner: if a cgroup is killed via cgroup.kill, it will fail to restart, because it must first be recreated. This problem will be addressed as well.

Logs

Nothing special, but comparing cgroups overviews can show the limits are lost after a service restart.

Environment

• Talos version: [talosctl version --nodes <problematic nodes>] e5acb10
• Kubernetes version: [kubectl version] v1.34.0
• Platform: QEMU