when restart kubelet, the gpu-device-plugin will Restart and Re-register to the new kubelet, leads to pods that depend on the gpu-device-plugin restarting as well.

[Soledao]

because of this, usually, these pods were not available for about 60 seconds.

Normally, a pod that uses only CPU and memory will not restart when restarting kubelet. What solutions can be taken to ensure that pods rely on gpu device are not unavailable when restarting kubelet? Especially when I want to upgrade kubelet.

[grac3gao-zz]

Hi @Soledao, could you provide the way to reproduce that (and the information about your cluster)?
I was trying to reproduce it by:

1. create a cluster with gpu -> install device installer pod to make sure device plugin starts working-> apply a pod using gpu
2. ssh into the node, and restart the kubelet sudo systemctl restart kubelet

However, I didn't see gpu-device-plugin get restart, the Pod uses gpu didn't restart either.