v1.8.1 loaded ARM image instead of AMD for rook/ceph:v1.15.5

[muhlba91]

Bug Report

Description

I updated Rook Ceph from v1.15.4 to v1.15.5 and the operator is not stuck in a CrashLoopBackOff due to exec /usr/local/bin/rook: exec format error. The operator was installed via the Helm chart.

Now, the image used is docker.io/rook/ceph@sha256:b94b23ecaf32e656c2460e2420dbc328be2951c673e41b7d3f2e3fe1eddecb59, which is a multi-arch image (the same as with v1.15.4).

I suppose that Talos loaded the "wrong" architecture onto the system and tries to run an ARM image on my Intel CPU.

Looking at the CLI commands, I couldn't find a way to delete an image, hoping it would download the correct architecture.

Can I re-trigger a download of an image? All other multi-arch images are working fine.

(FYI: I have also created a bug with rook but this does not seem to be related to rook itself as I seem to be the only one affected: rook/rook#14988)

Logs

N/A

Environment

• Talos version:

Server:
	NODE:        10.0.50.1
	Tag:         v1.8.1
	SHA:         477752fe
	Built:
	Go version:  go1.22.8
	OS/Arch:     linux/amd64
	Enabled:     RBAC

• Kubernetes version: v1.31.1
• Platform: Intel CPU

[smira]

Talos Linux doesn't pull the images either, it's done by the kubelet/CRI plugin.

I saw some similar issues, even though they seem to be quite random. You can try changing the image ref to a specific sha for the arch by using crane manifest on the multi-arch image.