To see the problem start a job inside a docker container and view the task/instance page. You'll cpu/ram/disk all at zero regardless of their actual usage.
I see errors like this in the thermos observer log:
This is likely because observer is running in a different pid namespace than the process. One solution would be for the runner to write out the pid namespace it is running in to the checkpoint and then have observer enter that namespace while sampling.
Or we can just get rid of the observer?