Re: Multivariate MCV stats can leak data to unprivileged users

On Sat, May 18, 2019 at 11:49:06AM -0700, Andres Freund wrote:
>Hi,
>
>On May 18, 2019 8:43:29 AM PDT, Dean Rasheed <dean(dot)a(dot)rasheed(at)gmail(dot)com> wrote:
>>On Sat, 18 May 2019 at 16:13, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>>
>>> Dean Rasheed <dean(dot)a(dot)rasheed(at)gmail(dot)com> writes:
>>> > On the other hand, pg_dump relies on pg_statistic_ext to work out
>>> > which extended statistics objects to dump. If we were to change
>>that
>>> > to use pg_stats_ext, then a user dumping a table with RLS using the
>>> > --enable-row-security flag wouldn't get any extended statistics
>>> > objects, which would be a somewhat surprising result.
>>>
>>> It seems like what we need here is to have a separation between the
>>> *definition* of a stats object (which is what pg_dump needs access
>>> to) and the current actual *data* in it. I'd have expected that
>>> keeping those in separate catalogs would be the thing to do, though
>>> perhaps it's too late for that.
>>>
>>
>>Yeah, with the benefit of hindsight, that would have made sense, but
>>that seems like a pretty big change to be attempting at this stage.

>
>Otoh, having a suboptimal catalog representation that we'll likely have
>to change in one of the next releases also isn't great. Seems likely
>that we'll need post beta1 catversion bumps anyway?
>

But that's not an issue intruduced by PG12, it works like that even for
the extended statistics introduced in PG10.

regards

--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services