From: Tom Lane <tgl@sss.pgh.pa.us>
To: Andreas Joseph Krogh <andreas@visena.com>
cc: pgsql-general@lists.postgresql.org
Subject: Re: Effects of REVOKE SELECT ON ALL TABLES IN SCHEMA pg_catalog FROM PUBLIC
In-reply-to: <VisenaEmail.0.81936c517b2d9cfe.191e44de951@origo-test01.app.internal.visena.net>
References: <VisenaEmail.0.81936c517b2d9cfe.191e44de951@origo-test01.app.internal.visena.net>
Comments: In-reply-to Andreas Joseph Krogh <andreas@visena.com>
	message dated "Thu, 12 Sep 2024 05:41:03 +0200"
MIME-Version: 1.0
Content-Type: text/plain; charset="UTF-8"
Content-ID: <3952714.1726115805.1@sss.pgh.pa.us>
Content-Transfer-Encoding: quoted-printable
Date: Thu, 12 Sep 2024 00:36:45 -0400
Message-ID: <3952715.1726115805@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/3952715.1726115805%40sss.pgh.pa.us>
Precedence: bulk

Andreas Joseph Krogh <andreas@visena.com> writes:
> Motivation: I have PowerBI users, with a separate =E2=80=98reporting=E2=80=
=99-role, accessing =

> a database and I want to prevent them from listing all tables, users, da=
tabases =

> and view-definitions (to not see the underlying query).

Postgres is not designed to support this requirement.

> I'm evaluating this:
> REVOKE SELECT ON ALL TABLES IN SCHEMA pg_catalog FROM PUBLIC; REVOKE SEL=
ECT ON =

> ALL TABLES IN SCHEMA information_schema FROM PUBLIC;
> Will this affect =E2=80=9Cnormal behaviour=E2=80=9D, ie. prevent the pla=
nner, or other =

> internal mechanisms, from working properly for sessions logged in with t=
he =

> =E2=80=98reporting=E2=80=99-role?

Probably 95% of that stuff will still work.  By the same token, there
are plenty of information-leaking code pathways that will still leak.
For instance, your restricted user will have no trouble discovering
the OIDs and names of all extant tables, using something like

do $$ begin
for tid in 1..1000000 loop
  if tid::regclass::text !=3D tid::text then
    raise notice 'tid % is %', tid, tid::regclass;
  end if; end loop;
end $$;

Functions such as pg_describe_object still work fine, too.

Experimenting with psql, a lot of stuff is broken as expected:

busted=3D> \d mytable
ERROR:  permission denied for table pg_class

but some things still work:

busted=3D> \sf sin
CREATE OR REPLACE FUNCTION pg_catalog.sin(double precision)
 RETURNS double precision
 LANGUAGE internal
 IMMUTABLE PARALLEL SAFE STRICT
AS $function$dsin$function$

This is pretty much the other side of the same coin.
The reason you can still parse and plan a query is that
it does not occur to large parts of the backend that there
should be any reason to refuse to read a system catalog.
That carries over to these operations as well.

This recent thread might be enlightening:

https://www.postgresql.org/message-id/flat/18604-04d64b68e981ced6%40postgr=
esql.org

If you have a requirement like this, I think the only safe
way to meet it is to not give those users direct SQL access.
Put some kind of restrictive app in front of the database.

			regards, tom lane