My guess is that we got a backoff greater than 12 with an error message and we ignored it. Then we used our default of 12 seconds (
here). Then we angered the API as we cleared the rest of our queue for the next several cycles of 12 seconds before falling back out of the backoff time we'd been given