On Wed, Aug 1, 2018 at 9:30 PM, Alexander Kukushkin <cyberdemn@gmail.com> wrote:
> Hello hackers,
>
> at the moment it is possible to reserve some amount of connection slots for
> superusers and this behavior is controlled by superuser_reserved_connections
> configuration parameter with the default value = 3.
>
> In case if all non-reserved connection slots are busy, replica fails to open
> a new connection and start streaming from the primary. Such behavior is very
> bad if you want to run postgresql HA clusters
Yeah, that's also bad if we want to use pg_baseback in the situation.
>
> Initially, replication connections required superuser privileges (in 9.0)
> and therefore they were deliberately excluded from
> superuser_reserved_connections. Basically that means it has never been
> possible to reserve come connection slots for replication connections.
>
> Later (9.1) it became possible to create a user with REPLICATION and
> NOSUPERUSER options, but comment in the postinit.c still tells that
> superuser is required.
>
> Now I think now it is a time to go further, and we should make it possible
> to reserve some connection slots for replication in a manner similar to
> superuser connections.
>
+1
> How should it work:
> 1. If we know that we got the replication connection, we just should make
> sure that there are at least superuser_reserved_connections free connection
> slots are available.
> 2. If we know that this is neither superuser nor replication connection, we
> should check that there are at least (superuser_reserved_connections +
> NumWalSenders() - max_wal_senders) connection slots are available.
You wanted to mean (superuser_reserved_connections + max_wal_senders -
NumWalSenders()) in the second point?
>
> And the last question how to control the number of reserved slots for
> replication. There are two options:
> 1. We can introduce a new GUC for that: replication_reserved_connections
> 2. Or we can just use the value of max_wal_senders
>
> Personally, I more like the second option.
>
One argrable point of the second option could be that it breaks
backward compatibility of the parameter configurations. That is, the
existing systems need to re-configure the max_connections. So it might
be better to take the first option with
replication_reservd_connections = 0 by default.
Regards,
--
Masahiko Sawada
NIPPON TELEGRAPH AND TELEPHONE CORPORATION
NTT Open Source Software Center