Re: pg_dump dump catalog ACLs

* Joe Conway (mail(at)joeconway(dot)com) wrote:
> On 03/01/2016 08:00 AM, Tom Lane wrote:
> > Joe Conway <mail(at)joeconway(dot)com> writes:
> >> Would it be a terrible idea to add some attribute to ACLs which can be
> >> used to indicate they should not be dumped (and supporting syntax)?
> >
> > Yes, we'd need some way to mark non-null ACLs as being "built-in
> > defaults". I do not see the need to have SQL syntax supporting that
> > though.
>
> I was thinking the supporting syntax might be used by extensions, for
> example.

I tend to agree with Tom that we don't really need SQL syntax for this.

> > Actually, wouldn't you need to mark individual aclitems as built-in
> > or not? Consider a situation where we have some function foo() that
> > by default has EXECUTE permission granted to some built-in "pg_admin"
> > role. If a given installation then also grants EXECUTE to "joe",
> > what you really want to have happen is for pg_dump to dump only the
> > grant to "joe". Mentioning pg_admin's grant would tie the dump to
> > a particular major PG version's idea of what the built-in roles are,
> > which is what I'm arguing we need to avoid.
>
> Yes, I guess it would need to be a per aclitem attribute.

Agreed.

> > I guess this could also be addressed by having two separate aclitem[]
> > columns, one that is expected to be frozen after initdb and one for
> > user-added grants.
>
> Yeah, that would work, but seems kind of ugly.

Rather than have two aclitem[] columns in every catalog, since this
information is only used by pg_dump and not during normal operation, we
could use the approach that pg_description took and have an independent
catalog table which just contains all non-NULL "system" ACLs. We could
populate it at the bottom of system_views.sql, so that we don't have to
explicitly think about updating that table whenever there's a change to
what the default ACLs are.

I don't see any reason it couldn't be used by extensions also, though
we'd have to do a bit more work on pg_dump to make it actually dump
out any non-default ACLs for extension-owned objects.

Thoughts?

Thanks!

Stephen