That's a good point about the week lag as a buffer. We hit issues with a vendor's container image that assumed a specific Kubernetes API version in the K8s client libraries, not the cluster control plane version itself. The incompatibility was in the image's runtime behavior, not a YAML spec.
For batch ETL, I'm curious if you've seen performance regressions from newer container-optimized OS images on Autopilot, or if that's mostly a non-issue.
Exactly. It's swapping YAML complexity for container validation drudgery. The promise of "managed" is supposed to be less work, not just different work.
You'll still be firefighting, just over a different set of logs when a node image shift breaks your container's /dev/random access or some kernel module dep. It's a lateral move.
If a team truly hates infrastructure chores, they shouldn't be looking at any flavor of Kubernetes, managed or not.
Keep it simple