Re: [EXTERNAL] Re: Support load balancing in libpq - Mailing list pgsql-hackers

From Michael Banck
Subject Re: [EXTERNAL] Re: Support load balancing in libpq
Date
Msg-id 63261554.df0a0220.44eb3.4f17@mx.google.com
Whole thread Raw
In response to Re: [EXTERNAL] Re: Support load balancing in libpq  (Maxim Orlov <orlovmg@gmail.com>)
Responses Re: Support load balancing in libpq
List pgsql-hackers
Hi,

On Wed, Sep 14, 2022 at 05:53:48PM +0300, Maxim Orlov wrote:
> > Also, IMO, the solution must have a fallback mechanism if the
> > standby/chosen host isn't reachable.
> 
> Yeah, I think it should. I'm not insisting on a particular name of options
> here, but in my view, the overall idea may be next:
> - we have two libpq's options: "load_balance_hosts" and "failover_timeout";
> - the "load_balance_hosts" should be "sequential" or "random";
> - the "failover_timeout" is a time period, within which, if connection to
> the server is not established, we switch to the next address or host.

Isn't this exactly what connect_timeout is providing? In my tests, it
worked exactly as I would expect it, i.e. after connect_timeout seconds,
libpq was re-shuffling and going for another host.

If you specify only one host (or all are down), you get an error.

In any case, I am not sure what failover has to do with it if we are
talking about initiating connections - usually failover is for already
established connections that suddendly go away for one reason or
another.

Or maybe I'm just not understanding where you're getting at?

> While writing this text, I start thinking that load balancing is a
> combination of two parameters above.

So I guess what you are saying is that if load_balance_hosts is set,
not setting connect_timeout would be a hazard, cause it would stall the
connection attempt even though other hosts would be available.

That's right, but I guess it's already a hazard if you put multiple
hosts in there, and the connection is not immediately failed (because
the host doesn't exist or it rejects the connection) but stalled by a
firewall for one host, while other hosts later on in the list would be
happy to accept connections.

So maybe this is something to think about, but just changing the
defaul of connect_timeout to something else when load balancing is on
would be very surprising. In any case, I don't think this absolutely
needs to be addressed by the initial feature, it could be expanded upon
later on if needed, the feature is useful on its own, along with
connect_timeout.


Michael



pgsql-hackers by date:

Previous
From: Alvaro Herrera
Date:
Subject: Re: missing indexes in indexlist with partitioned tables
Next
From: Tom Lane
Date:
Subject: Re: missing indexes in indexlist with partitioned tables