In analysis of parent issue we see that part of the problem is FLUSH starts fail under load as some participants take a bit longer than the default 2000ms start_flush_timeout to reply. 2000ms is quite short. We'll increase this to 10000 ms.