Re: connections

> On Sun, Apr 20, 2008 at 6:00 PM, Dave Page <dpage(at)pgadmin(dot)org> wrote:
> > On Fri, Apr 18, 2008 at 7:30 PM, Roberts, Jon
<Jon(dot)Roberts(at)asurion(dot)com>
> wrote:
> >
> > I'm awaiting confirmation from the lead architect, but I believe
that
> > a single connection from pgAdmin will result in 1 connection to each
> > node in the cluster in GridSQL. However I believe there is also a
> > certain amount of pooling going on so I don't believe it's such a
> > problem. Besides.... that may be 16 connections in a 16 node system,
> > but each of those is a complete PostgreSQL/EnterpriseDB server
capable
> > of handling as many connections as it could if it were a standalone
> > server (which could be hundreds or thousands), so it's no more an
> > issue than if you had a single server.
>
> The word from the GridSQL guru is:
>
> I am not sure I understand why he has so many connections because each
> segment should be on a different host. Maybe they have something set
> up where there are multiple logical nodes (er, segments) running on a
> single postmaster instance.
>

They run a postmaster instance per core so we have 4 databases running
per segment host. We have four segment hosts so that means we have 16
databases running. Then there is one more database which is the
"master" which is equivalent to the GridSQL Coordinator.

> Maybe that further explodes if each segment creates a connection to
> every other one and persists it.
>

I'm just a customer so I don't know for sure but I'm guessing it does.
They recommend each segment database have max_connections set to 5-10
higher than the master and they have a large interconnect network
between the servers.

> In GridSQL, we have a pool of connections that are used, starting with
> 5 for each underlying node. These are shared amongst the users, unless
> somethings stateful happens in which case the underlying connection
> needs to become persistent for a user session. We also have an extra
> connection for the metadata database, and some vestiges of old code
> for a pool of "coordinator" connections, but they are not used for
> most queries.
>
> Anyway, the bottom line is, GridSQL will not consume as many
> connections as in the Greenplum case.

Dave, thanks for the detailed response. I really do appreciate it.

Jon