Re: Mystery with REVOKE PRIVILEGE

Konstantin Knizhnik <knizhnik(at)garret(dot)ru> writes:
>> The problem really seems to be in `select_best_grantor` - it choose
>> "wrong" grantor.
>> It seems to be the bug because as a result of this operation no
>> privilege is revoked at all (because then `merge_acl_with_grant` found
>> no match).
>> I wonder if `select_best_grantor` should always prefer exact match?

> I mean something like this (see attached patch).

I don't think "let's make select_best_grantor even more magic"
is the right approach. IMO, if there is a GRANTED BY clause,
we should use exactly that grantor and not apply select_best_grantor
at all. This is, for example, certainly the behavior needed for
pg_dump.

If I'm reading the SQL spec correctly, it only allows "GRANTED BY
CURRENT_ROLE" and "GRANTED BY CURRENT_USER", so we're already
extending the spec by accepting "GRANTED BY somebody". But it's
really hard to find any justification for select_best_grantor in
the spec. I think we ought to redefine that as a legacy behavior
we use in the absence of any explicit GRANTED BY clause.

regards, tom lane