Thread: Why is vacuum_defer_cleanup_age PGC_USERSET?

It seems to me it ought to be PGC_SIGHUP.  In the first place, the
actually important value under normal circumstances is going to be
that seen by autovac workers.  There is also some impact on HOT
cleanup, but I can't see any sane argument why you don't want all
backends to be applying the same value for that, since it's difficult
or impossible to predict which session is actually going to HOT-clean
any page.  In the second place, it seems completely foolish to let
ordinary unprivileged users mess with the value.  Arguably an individual
user could produce a denial of service on HS slaves by cutting his local
value of vacuum_defer_cleanup_age and then vacuuming tables that queries
on the slaves will look at.  To the extent that the variable does
anything useful at all, it is the system-wide behavior that is
important, so I see no use-case for changing it in individual sessions
anyway.

It also appears to me to be misclassified.  WAL_STANDBY_SERVERS should
be the category for variables that you'd adjust on an HS slave, no?
But this is something that has to be set on the master.  Possibly the
best place for it is WAL_REPLICATION.
        regards, tom lane

On Sat, 2010-07-03 at 13:07 -0400, Tom Lane wrote:
> It seems to me it ought to be PGC_SIGHUP.  In the first place, the
> actually important value under normal circumstances is going to be
> that seen by autovac workers.  There is also some impact on HOT
> cleanup, but I can't see any sane argument why you don't want all
> backends to be applying the same value for that, since it's difficult
> or impossible to predict which session is actually going to HOT-clean
> any page.  In the second place, it seems completely foolish to let
> ordinary unprivileged users mess with the value.  Arguably an individual
> user could produce a denial of service on HS slaves by cutting his local
> value of vacuum_defer_cleanup_age and then vacuuming tables that queries
> on the slaves will look at.  To the extent that the variable does
> anything useful at all, it is the system-wide behavior that is
> important, so I see no use-case for changing it in individual sessions
> anyway.

Happy with that argument, so agreed.

> It also appears to me to be misclassified.  WAL_STANDBY_SERVERS should
> be the category for variables that you'd adjust on an HS slave, no?
> But this is something that has to be set on the master.  Possibly the
> best place for it is WAL_REPLICATION.

Those didn't exist when it was originally classified, thats all.

Yes, those categories sounds good.

-- Simon Riggs           www.2ndQuadrant.com

On 7/3/10 10:07 AM, Tom Lane wrote:
> It seems to me it ought to be PGC_SIGHUP.

+1

I believe it's userset because the other vacuum parameters are, and that
nobody thought about it much.


--                                  -- Josh Berkus                                    PostgreSQL Experts Inc.
                        http://www.pgexperts.com