We are testing this in production right now as the symptoms he describes are something I’ve been trying to track down for almost a year. It appears like the be our issue as I was unable to see the problem on our machines where /tmp is using tmpfs.