2012/6/26 Tom Lane <tgl@sss.pgh.pa.us>:
> Kohei KaiGai <kaigai@kaigai.gr.jp> writes:
>> 2012/6/26 Robert Haas <robertmhaas@gmail.com>:
>>> I think you're missing the point. Everyone who has commented on this
>>> issue is in favor of having some check that causes the RLS predicate
>>> *not to get added in the first place*.
>
>> Here is a simple idea to avoid the second problematic scenario; that
>> assign 0 as cost of has_superuser_privilege().
>
> I am not sure which part of "this isn't safe" isn't getting through to
> you. Aside from the scenarios Robert mentioned, consider the
> possibility that f_malicious() is marked immutable, so that the planner
> is likely to call it (to replace the call with its value) before it will
> ever think about whether has_superuser_privilege should be called first.
>
> Please just do what everybody is asking for, and create a bypass that
> does not require fragile, easily-broken-by-future-changes assumptions
> about what the planner will do with a WHERE clause.
>
The problem is the way to implement it.
If we would have permission checks on planner stage, it cannot handle
a case when user-id would be switched prior to executor stage, thus
it needs something remedy to handle the scenario correctly.
Instead of a unique plan per query, it might be a solution to generate
multiple plans depending on user-id, and choose a proper one in
executor stage.
Which type of implementation is what everybody is asking for?
Thanks,
--
KaiGai Kohei <kaigai@kaigai.gr.jp>