Re: Weird pg_dumpall bug?

I did a dump of a 7.4.11 database using the 8.1.2 pg_dumpall. I got
this at the top of the dump:

...
...
CREATE ROLE support;
ALTER ROLE support WITH NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB
LOGIN PASSWORD 'md5xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx';
...
...
CREATE ROLE support;
ALTER ROLE support WITH NOSUPERUSER INHERIT NOCREATEROLE NOCREATEDB NOLOGIN;
...
...

It dumped the "support" role twice!

Any ideas?

Hmmmm...actually. It's because I have a user called 'support' and a
group called 'support'.

Seems like it needs a fix...

Chris

* Christopher Kings-Lynne (chriskl(at)familyhealth(dot)com(dot)au) wrote:
> Hmmmm...actually. It's because I have a user called 'support' and a
> group called 'support'.
>
> Seems like it needs a fix...

Have you got a suggestion on just how to fix it...? Debian's
pg_upgradecluster bails out with an error when it discovers this
situation but I don't think it'd be sensible for pg_dump to do that...

Thanks,

Stephen

On Tue, 2006-01-24 at 09:44 -0500, Stephen Frost wrote:
> * Christopher Kings-Lynne (chriskl(at)familyhealth(dot)com(dot)au) wrote:
> > Hmmmm...actually. It's because I have a user called 'support' and a
> > group called 'support'.
> >
> > Seems like it needs a fix...
>
> Have you got a suggestion on just how to fix it...? Debian's
> pg_upgradecluster bails out with an error when it discovers this
> situation but I don't think it'd be sensible for pg_dump to do that...
>

How about an option to map groups whose names conflict with user names
using a prefix mechanism?

e.g. --map-conflicting-groups=gr_

Then in Christopher's example his support group would become the role
gr_support.

Just a thought.

cheers

andrew

Christopher Kings-Lynne <chriskl(at)familyhealth(dot)com(dot)au> writes:
> Hmmmm...actually. It's because I have a user called 'support' and a
> group called 'support'.

This was discussed already. It's a corner case we didn't really think
about while designing roles. It's possible to support this: the group
and the user will now really be the same entity, ie a role that has both
its own login privileges and members. However pg_dump isn't doing the
right things to make the old situation morph into the new one. Want to
write a patch?

regards, tom lane

On Tue, 2006-01-24 at 10:05 -0500, Tom Lane wrote:
> Christopher Kings-Lynne <chriskl(at)familyhealth(dot)com(dot)au> writes:
> > Hmmmm...actually. It's because I have a user called 'support' and a
> > group called 'support'.
>
> It's possible to support this: the group
> and the user will now really be the same entity, ie a role that has both
> its own login privileges and members.
>

Assuming you actually want to unify the two objects. That might well be
the common case, but will it always be true?

cheers

andrew

Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> On Tue, 2006-01-24 at 10:05 -0500, Tom Lane wrote:
>> It's possible to support this: the group
>> and the user will now really be the same entity, ie a role that has both
>> its own login privileges and members.

> Assuming you actually want to unify the two objects. That might well be
> the common case, but will it always be true?

As compared to what? I didn't like the notion of auto-renaming one of
the roles, if that's what you're suggesting. That seems well outside
pg_dump's charter.

regards, tom lane

Am Dienstag, 24. Januar 2006 15:44 schrieb Stephen Frost:
> Have you got a suggestion on just how to fix it...? Debian's
> pg_upgradecluster bails out with an error when it discovers this
> situation but I don't think it'd be sensible for pg_dump to do that...

Why not? If the backup cannot be made in a way such that the semantics of the
restored database are the same, it shouldn't do it.

--
Peter Eisentraut
http://developer.postgresql.org/~petere/

Peter Eisentraut <peter_e(at)gmx(dot)net> writes:
> Am Dienstag, 24. Januar 2006 15:44 schrieb Stephen Frost:
>> Have you got a suggestion on just how to fix it...? Debian's
>> pg_upgradecluster bails out with an error when it discovers this
>> situation but I don't think it'd be sensible for pg_dump to do that...

> Why not? If the backup cannot be made in a way such that the
> semantics of the restored database are the same, it shouldn't do it.

If you take a hard line on that position, then it's not necessary for
pg_dump to support cross-version operation at all, because no major
PG release is ever 100.0% compatible with the previous one.

What is actually required of pg_dump is that it produce the closest
approximation it can get to the old behavior within the context of the
new version's semantics. I can easily cite half a dozen examples of
cases where we've applied this logic in previous versions. I do not
see a reason to treat this case differently. The difference between
a single role acting as both user and group and the prior behavior of
separate objects is certainly well within the "fuzz factor" that we've
allowed pg_dump to have in the past.

regards, tom lane

> How about an option to map groups whose names conflict with user names
> using a prefix mechanism?
>
> e.g. --map-conflicting-groups=gr_
>
> Then in Christopher's example his support group would become the role
> gr_support.

No bad, have to change some application code then as well...

Chris

On Tue, Jan 24, 2006 at 10:42:17AM -0500, Tom Lane wrote:
> Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> > On Tue, 2006-01-24 at 10:05 -0500, Tom Lane wrote:
> >> It's possible to support this: the group
> >> and the user will now really be the same entity, ie a role that has both
> >> its own login privileges and members.
>
> > Assuming you actually want to unify the two objects. That might well be
> > the common case, but will it always be true?
>
> As compared to what? I didn't like the notion of auto-renaming one of
> the roles, if that's what you're suggesting. That seems well outside
> pg_dump's charter.

If you want something renamed, you can handle that case by just renaming
it before you do the dump, but it would be nice if pg_dump would raise a
nice big warning when this condition exists so you're aware of it. Or
maybe even refuse to run unless you supply some command line option to
over-ride.

I don't think we should morph the two together by default either,
because that's very possibly not what the user originally intended.
--
Jim C. Nasby, Sr. Engineering Consultant jnasby(at)pervasive(dot)com
Pervasive Software http://pervasive.com work: 512-231-6117
vcard: http://jim.nasby.net/pervasive.vcf cell: 512-569-9461