Since we launch the task after pivot_root, we no longer has the access to the mesos-health-check binary. The solution is to refactor health check to be a library (libprocess) so that it does not depend on the underlying filesystem.
One note here is that we should strive to keep both the command executor and the task in the same mount namespace so that Mesos CLI tooling does not need to find the mount namespace for the task. It just need to find the corresponding pid for the executor. This statement is arguable, see the comment below.