Re: replacing role-level NOINHERIT with a grant-level option

On Thu, Jun 02, 2022 at 03:37:34PM -0400, Robert Haas wrote:
> On Thu, Jun 2, 2022 at 2:07 PM Nathan Bossart <nathandbossart@gmail.com> wrote:
>> I think we should also consider replacing role attributes with predefined
>> roles.  I'm not sure that this proposal totally prepares us for such a
>> change, given role attributes apply only to the specific role for which
>> they are set and aren't inherited.  ISTM in order to support that, we'd
>> need even more enhanced functionality.  For example, if I want 'robert' to
>> be a superuser, and I want 'joe' to inherit the privileges of 'robert' but
>> not 'pg_superuser', you'd need some way to specify inheriting only certain
>> privileges possessed by an intermediate role.
> 
> I guess we could think about adding something like an ONLY clause,
> like GRANT ONLY robert TO joe. I feel a little bit uncomfortable about
> that, though, because it assumes that robert is a superuser but his
> own privileges are distinguishable from those of the superuser. Are
> they really? If I can assume robert's identity, I can presumably
> Trojan my way into the superuser account pretty easily. I'll just
> define a little trigger on one of his tables. I don't really see a way
> where we can ever make it safe to grant a non-superuser membership in
> a superuser role.

I was primarily looking at this from the angle of preserving current
behavior when upgrading from a version with role attributes to a version
without them.  If it's alright that a role with privileges of a superuser
role begins being treated like a superuser after an upgrade, then we
probably don't need something like GRANT ONLY.  I bet that's how a lot of
people expect role attributes to work, anyway.  I'm sure I did at some
point.

> But even if there is a way, I think that is a separate patch from what
> I'm proposing here. [NO]INHERIT only has to do with what privileges
> you can exercise without SET ROLE. To solve the problem you're talking
> about here, you'd need a way to control what privileges are conferred
> in any manner, which is related, but different.

I agree that the role-attribute-to-predefined-role stuff needs its own
thread.  I just think it's worth designing this stuff with that in mind.

-- 
Nathan Bossart
Amazon Web Services: https://aws.amazon.com