Description of problem:
When a pod uses a PersistentVolumeClaim, this can happen:



Version-Release number of selected component (if applicable):
origin-3.1.0

How reproducible:
~30%-50%

Steps to Reproduce:
1. create a PV and a claim (I use Cinder volumes, but I saw it on AWS and GCE too)
2. create a pod that uses the claim
3. In a loop:
  3.1 create the pod
  3.2 wait until it's running
  3.3 run 'kubectl describe pods'
  3.4 delete it
  3.5 wait until the volume is unmounted and detached from the node (this is important!)

Actual results:
at step 3.3, you can see it was necessary to restart the pod container(s), "kubectl describe pods" shows something like:

Error syncing pod, skipping: not all containers have started: 0 != 1

Expected results:
The containers start on the first try.

Additional info:
The pod is started eventually, it's just slower (~1 minute in my OpenStack setup, attaching a volume is slow...).

Fix: https://github.com/kubernetes/kubernetes/pull/19600