Rolling Update support inplace update pod image

[xulinfei1996]

What you would like to be added?

Rolling update supports update image while not delete pod just update pod(InPlaceIfPossible), so that kubelet will download image and restart container.

Why is this needed?

Currently, kubelet supports downloading the image and restarting the container without the deletion of the entire Pod to update the image.
When the cluster is busy, deleting a Pod and rescheduling it (may be PodGroup), incurs significant overhead (e.g., resource reallocation, scheduling delays, etc.). Therefore, avoiding Pod deletion and directly updating the image can significantly reduce this overhead.

[gflarity]

Hi @xulinfei1996, this is an interesting idea, but this approach is generally discouraged even though it works. Can you give a more concrete example of why for AI workloads in particular this overhead is bad and something like this would be needed?