The test was flaky because it required the job succeeds 3 times with
pseudorandom 50% failure chance within 15 minutes, while there is an
exponential back-off delay (10s, 20s, 40s …) capped at 6 minutes before
recreating failed pods. As 7 consecutive failures (1/128 chance) could
take 20+ minutes, exceeding the timeout, the test failed intermittently
because of "timed out waiting for the condition".
This PR forces the Pods of a Job to be scheduled to a single node and
uses a hostPath volume instead of an emptyDir to persist data across new
Pods.
