The kubelet logic for figuring out which docker container to use to show logs from is flawed. It's currently returning a container uuid to actually look at somewhat randomly (especially when many attempts occur in a row because the pod is restarting).

A short term fix would be to use the most recent attempt. In the long term, it would be more appropriate for the logs to be across attempts (perhaps delimited by the attempt id) with a more sophisticated set of parameters available for consumers to clarify which logs exactly they need.