Re: Row-level Security vs Application-level authz

On 2/24/15 3:31 PM, Stephen Frost wrote:
> * David Steele (david(at)pgmasters(dot)net) wrote:
>> On 2/24/15 3:07 PM, Stephen Frost wrote:
>>> The problem with a temporary table is, well, it goes away. :) There are
>>> further concerns that, because it's created in some fashion by the
>>> single application user, it might be less secure. Really, though, I'd
>>> want it to be real so that it could have constraints be on it which
>>> reference other appropriate tables, so the web user doesn't have to have
>>> rights in any fashion to create objects, and so that it can be
>>> referenced from RLS policies. A table as transient as a temporary table
>>> doesn't strike me as the right solution for that.
>>
>> Temp tables go away at the end of the session, sure. It seems like
>> exactly the time when you'd want them to do so.
>>
>> If the temp table is created by a security definer function (as was
>> suggested earlier) then no special user privs are required.
>>
>> Being referenced from RLS polices is a good argument, though. I guess
>> that's not possible with a temp table? Are they pre-parsed?
>
> Actually, it is possible, but it creates a dependency on the temporary
> table and when the temporary table goes away, so will the policy. This
> isn't a huge issue for RLS, of course, as if there's no policy then the
> default-deny policy will kick in, but you can't have a policy against a
> temporary table continue past the end of that session.

It's not good if policies go away, even with a default.

> Following the earlier discussion, I suppose you could create both the
> temporary table and the policies in the security definer function, but
> it feels a lot cleaner to have a real table for all of that, in my view,
> to keep that security definer function nice and simple.

Then you'd have a policy for each user which sounds messy, or some crazy
locking scheme which sounds worse. I agree that a real table sounds
like a better solution.

> Further, there's lots of other reasons to have a session table anyway,
> from an application standpoint, and so this feels like an approach which
> is more in-line with how the application likely wants to operate anyway.
> It's also handy to be able to log into the database and see all the
> current sessions, similar to how we have pg_stat_activity.

Well, I am a fan of monitoring.

So I guess my last question is if you are inserting rows into a table to
track user connections, how do you clean them out when the client does
not disconnect cleanly? Or is this table intended to be append-only?

--
- David Steele
david(at)pgmasters(dot)net