=?ISO-8859-1?Q?G=E1bor_Farkas?= <gabor@nekomancer.net> writes:
> and this is still a mystery for me, because i understand that
> idle-in-transaction is wrong, but even so, a process that i start today,
> in my opinion simply cannot block the recovery of dead rows, that were
> deleted yesterday.
Well, it's not one but two longest-transaction-lifetimes. That is,
take the oldest transaction that's running now, and consider the
oldest transaction that was running when it started. VACUUM can
reclaim rows that were deleted by transactions that started before
that one. You didn't explain your application's behavior exactly,
but does that help?
regards, tom lane