On Thu, Apr 1, 2010 at 9:09 AM, Heikki Linnakangas
<heikki.linnakangas@enterprisedb.com> wrote:
> Fujii Masao wrote:
>> On Thu, Apr 1, 2010 at 11:00 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>>> On Wed, Mar 31, 2010 at 9:58 PM, Fujii Masao <masao.fujii@gmail.com> wrote:
>>>> You mean that we should change replication connection not to consume
>>>> superuser_reserved_connections slots in 9.0?
>>> Yes.
>
> I think it's good that walsenders can use the superuser reserved slots,
> that way a client that opens max_connections connections can't block out
> standby servers from connecting.
>
>> Preventing superuser connections from consuming superuser_reserved_connections
>> slots seems strange for me. So I'm leaning toward just removing superuser
>> privilege from replication connection again. Thought?
>
> That would be good, but I fear it's a bigger change than we should be
> doing at this point.
>
> How about we adjust the backends math a bit:
>
> Currently:
>
> ReservedBackends = superuser_reserved_connections
> MaxBackends = max_connections + autovacuum_max_workers + 1;
>
> Proposal:
>
> ReservedBackends = superuser_reserved_connections + max_wal_senders
> MaxBackends = max_connections + autovacuum_max_workers + max_wal_senders + 1
>
> So we implicitly reserve a slot and a superuser reserved slot for each
> walsender. Walsenders use the slots reserved for superusers, but if you
> set superuser_reserved_connections=3, there's still always at least
> three slots available for superuser to log in with psql, even if the
> maximum number of walsenders are connected.
That seems pretty reasonable to me. I haven't checked how much code
impact there is. I know Tom doesn't think we should change it at all,
but surely pre-beta is the time to fix nasty corner cases that were
added by recently committed patches?
...Robert
