August Zajonc <augustz@augustz.com> writes:
> The key is how lightweight the setup could be, which matters because
> clients are not always willing to pay for a PITR setup. The low overhead
> would mean you'd feel fine about setting guc to 1hr or so.
This would have exactly the same performance consequences as always
having an hour-old open transaction. I'm afraid that describing it
as "low overhead" is mere wishful thinking: it would cripple vacuuming
of high-update tables and greatly increase the typical load on pg_clog
and pg_subtrans. We already know that pg_subtrans contention can be a
source of context-swap storms, with the size of the window back to
GlobalXmin being the controlling factor for how bad it gets.
It's possible that this last could be addressed by separating the
concept of "old enough to be vacuumed" from GlobalXmin, but it's
certainly not a trivial thing.
regards, tom lane
