Re: BUG #6689: socket file name convention doesn't allow different IPs and the same port for different pgclusters

On Fri, Jun 15, 2012 at 7:33 PM, Mark Schneider
<ms(at)it-infrastrukturen(dot)org> wrote:
> Am 15.06.2012 07:56, schrieb Magnus Hagander:
>
>> On Wed, Jun 13, 2012 at 2:45 AM,<ms(at)it-infrastrukturen(dot)org>  wrote:
>>>
>>> The following bug has been logged on the website:
>>>
>>> Bug reference:      6689
>>> Logged by:          Mark
>>> Email address:      ms(at)it-infrastrukturen(dot)org
>>> PostgreSQL version: 9.1.3
>>> Operating system:   unbuntu-server 12.04 LTS
>>> Description:
>>>
>>> When using /var/run/postgresql/ as unix_socket_directory for "main" and
>>> "second" pgcluster and different IP-interfaces (addresse) for every
>>> pgcluster (but the same default port number), there are different pid
>>> file
>>> names (like 9.1-main.pid and 9.1-secondpg.pid) *but* names of unix_sockek
>>> files doesn't follow such name conventions.
>>>
>>> It results in error when trying to run the second pgcluster with another
>>> IP
>>> but the same (default port 5432):
>>> Error: Port conflict: another instance is already running on on
>>> /var/run/postgresql with port 5432
>>
>> First of all, this is not a bug - this is intended behaviour. Perhaps
>> the documentation needs to be clearified on the fact that the port
>> number needs to be unique across all instances though.
>
> Even it is not an implementation bug it is a kind of *conception* bug from
> my point of view.

That's why I'm suggesting updating the documentation to make that more clear.

> There is *no* real reason why port numbers should be different for different
> IP addresses.

True. But as I said, it doesn't just control the TCP port *or* the
unix socket name.

> There is anything about running further psql instances on different ethernet
> interfaces (on one box) in the current doc yet.

Right. That's my suggested fix.

>> One reason is that the port number is used to control several other
>> things. Just the Unix socket is only one of the things - it also
>> controls the name of shared memory segments, so it still won't work if
>> you work around that.
>
> When I want to run two separate psql instances they should use also separate
> shared memory segments.

Exaclty. Which is why they need separate port numbers. If they use the
same port number, they'd use the same shared memory segment...

>> All instances need to have their own ports. If you want to listen on
>> the same port on different IPs, you will need to use something like
>> kernel level ip forwarding to rewrite the access, and actually run
>> PostgreSQL on different ports still.
>
> Using different port number for the second ethernet interface works as
> expected.

Yes. For just the ethernet interface. But again, the parameter
controls more things.

> How complex is it to implement different unix_socket names in the sourse
> code?
> (when further instances listen on different ethernet interfaces)

Probably not hugely so, but it wouldn't fix your problem.

There's an ongoing discussoin on-hackers about being able to listen to
multiple unix sockets, and multiple port numbers. But as you'll see if
you read up on that one (see archives.postgresql.org), it's all still
requiring a single "primary port" to control the other parts. So that
also won't solve your problem.

--
 Magnus Hagander
 Me: http://www.hagander.net/
 Work: http://www.redpill-linpro.com/