This job is too flaky to be of much value right now. I propose to allow it to fail until we can observe more data and then set the max and margin appropriately.