Thread: NOTIFY channel

Hi,

I'm curious what the design assumptions of the LISTEN / NOTIFY are.  For example, would it work to have a thousand
LISTENchannels open at the same time or would that be abusing the assumptions that were made?  In short, I'm trying to
understandif I need a few central LISTEN / NOTIFY channels that then distribute the events out or not. 

Thank you for your time,
Perry

On Fri, Apr 26, 2013 at 11:30 AM, Perry Smith <pedzsan@gmail.com> wrote:
> Hi,
>
> I'm curious what the design assumptions of the LISTEN / NOTIFY are.  For example, would it work to have a thousand
LISTENchannels open at the same time or would that be abusing the assumptions that were made?  In short, I'm trying to
understandif I need a few central LISTEN / NOTIFY channels that then distribute the events out or not. 

Couple quick points about notify:
1) In somewhat recent postgres (9.0) notify implementation was
significantly optimized and expanded to include a payload.

2) notify does not work with pgbouncer (yet), although I have in the
past had a private pgbouncer with functioning notify.

3) notifies are not delivered until end of transaction, which is
sometimes useful, sometimes annoying

4) delivery of notifications is 'best effort', but that effort is
pretty much always successful.

5) listener's in the range of 1000's is ok.  but don't be afraid to
think about using payload and/or putting specific instructions into a
table that the client listener acts on after receiving notify.

merlin

Perry Smith <pedzsan@gmail.com> writes:
> I'm curious what the design assumptions of the LISTEN / NOTIFY are.  For example, would it work to have a thousand
LISTENchannels open at the same time or would that be abusing the assumptions that were made?  In short, I'm trying to
understandif I need a few central LISTEN / NOTIFY channels that then distribute the events out or not. 

In the current implementation (since 9.0 IIRC), I don't believe the
total number of channels in use affects the transport mechanism at all.
However, the number that a particular backend is listening on would
affect how fast it can process the messages --- we just run through the
local LISTEN-names list serially for a match to see if an incoming
message is interesting or not.  So I'd try to avoid having individual
sessions listening to very large numbers of channels.  OTOH, the serial
list scan might still be faster than anything you could easily do in SQL
or plpgsql to sort through messages you were trying to route manually.

            regards, tom lane