Magnus Hagander <magnus@hagander.net> writes:
> On Thu, Nov 15, 2012 at 6:53 PM, Peter Eisentraut <peter_e@gmx.net> wrote:
>> The only thing you couldn't handle that way are SIGHUP settings, but the
>> often-cited use cases work_mem, logging, etc. would work.
> How hard would it be to make it work for SIGHUP?
One issue is that pg_db_role_setting is currently considered only at
session start, and unless you want to hack that somehow, these new
settings would only be absorbed by freshly-started sessions.
Now, there's already a good deal of asynchrony in when individual
processes notice postgresql.conf updates, whether they're for SIGHUP
or lesser settings. So maybe that's all right. If you weren't happy
about it, one of several things you'd have to work out is how to send a
SIGHUP only after you've committed the changes.
Another and probably bigger thing is that SIGHUP is used for settings
that do something useful only in background processes (eg checkpointer).
Some of those processes are not capable of reading system catalogs at
all. This is particularly a showstopper for settings affecting the
postmaster itself, which is most certainly *not* going to grow the
ability to read catalogs.
On the whole I suspect the existing push towards rewritable config file
entries is going to go further in less time for anything whose effects
aren't limited to regular backend sessions. I don't object to Peter's
idea as such, but it's not going to help us for SIGHUP settings.
regards, tom lane
