Re: SET ROLE and reserved roles

On Fri, Apr 15, 2016 at 11:56 AM, Stephen Frost <sfrost(at)snowman(dot)net> wrote:
> Perhaps it would be helpful to return to the initial goal of these
> default roles.
>
> We've identified certain privileges which are currently superuser-only
> and we would like to be able to be GRANT those to non-superuser roles.
> Where our GRANT system is able to provide the granularity desired, we
> have simply removed the superuser checks, set up appropriate default
> values and, through the "pg_dump dump catalog ACLs" patch, allowed
> administrators to make the changes to those objects via the
> 'GRANT privilege ON object TO role' system.
>
> For those cases where our GRANT system is unable to provide the
> granularity desired and it isn't sensible to extend the GRANT system to
> cover the exact granularity we wish to provide, we needed to come up
> with an alternative approach. That approach is the use of special
> default roles, where we can allow exactly the level of granularity
> desired and give administrators a way to grant those privileges to
> users.
>
> What this means in a nutshell is that we've collectivly decided that
> we'd rather support:
>
> /* uses the 'GRANT role TO role' system */
> GRANT pg_signal_backend TO test_user;
>
> than:
>
> /*
> * uses the 'GRANT privilege ON object TO role' system
> * seems like cluster-level is actually the right answer here, but
> * we don't support such an object type currently, so using database
> * for the example
> */
> GRANT SIGNAL BACKEND ON DATABASE my_database TO test_user;
>
> The goal being that the result of the two commands is identical (where
> the second command is only hypothetical at this point, but hopefully the
> meaning is conveyed).

I quite understand all of that. The problem isn't that I don't
understand what you did. The problem is that I don't agree with it.
I don't think that the way you implemented is a good idea, nor do I
think it reflects the consensus that was reached on list. There was
agreement to create some special magic roles. There was not agreement
around the special magic you've sprinkled on those roles that makes
them work unlike all other roles in the system, and I think that
special magic is a bad idea.

> However, GRANT'ing a role to a user traditionally also allows the user
> to 'SET ROLE' to that user, to create objects as that user, and other
> privileges. This results in two issues, as I see it:
>
> 1) The administrator who is looking to grant the specific 'signal
> backend' privilege to a user is unlikely to intend or desire for that
> user to also be able to SET ROLE to the role, or for that user to be
> able to create objects as the default role.

I don't think being able to SET ROLE to that role is something that we
should be trying to prevent. You're already discovering why. You
tried to create this new special kind of role that you can't become,
and there are already various reports of cases that you've missed.
You will keep finding more for a while, I predict. If that role is
minimally privileged, who cares if users can become it?

> 2) If the default role ends up owning objects then we have much less
> flexibility in making changes to that role because we must ensure
> that those objects are preserved across upgrades, etc.

Tom already said his piece on this point, and I think he's right. You
re-stating the argument doesn't change that. In any case, if we want
to prevent this, I bet there's some less buggy and invasive way in
which it could be done than what you've actually chosen.

> Further, there seems to be no particular use-case which is satisfied by
> allowing SET ROLE'ing to the default roles or for those roles to own
> objects; indeed, it seems more likely to simply spark confusion and
> ultimately may prevent administrators from making use of this system for
> granting privileges as they are unable to GRANT just the specific
> privilege they wish to.

Great. But there's no particular use case served by a lot of things
which are natural outgrowths of the rest of the system which we permit
anyway because it's too awkward otherwise - like zero-column tables.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company