On Fri, Jul 9, 2010 at 1:21 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Simon Riggs <simon@2ndQuadrant.com> writes:
>> On Fri, 2010-07-09 at 11:09 -0400, Tom Lane wrote:
>>> Simon Riggs <simon@2ndQuadrant.com> writes:
>>>> Strangely, I was looking into removing the ExecCheckRTPerms check
>>>> altogether by forcing plan invalidation when permissions are updated.
>>>> That would be a performance tweak that would render this change useless.
>>>
>>> That seems both pointless and wrong. Permissions checks should happen
>>> at execution time not plan time.
>
>> Agreed that permission checks should logically be applied at execution
>> time. I am proposing a performance optimisation, not a change in
>> behaviour.
>
> Except that it *is* a change in behavior: the first check will occur too
> soon.
You might be able to get around this by doing the first check on first
use of the plan and then going and marking all the plans as needing a
recheck whenever a permissions change happens. Whether the
performance savings are sufficient to justify such a thing is another
matter.
> The fact that we're interested in adding plugin permissions checking
> pretty much destroys the idea anyway. You cannot assume that a plan
> cache invalidation will happen for any change in external state that
> a plugin might be consulting.
This is certainly true, but I also wonder what SE-PostgreSQL plans to
do about this. Taking this to its logical exteme, the system security
policy could change in mid-query - and while you'd like to think that
the system would stop emitting tuples on a dime, that's probably not
too feasible in practice. I am assuming that SE-PostgreSQL will want
to do some kind of caching, but I wonder how one decides what to cache
and for how long, and whether there's any mechanism for propagating
cache invalidations.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company