Re: [HACKERS] Improving RLS planning

On 29 December 2016 at 15:55, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Dean Rasheed <dean.a.rasheed@gmail.com> writes:
>> On 28 December 2016 at 19:12, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> [ getting back to this patch finally... ]  I made the suggested change
>>> to that test case, and what I see is a whole lot of "NOTICE: snooped value
>>> = whatever" outputs.
>>>
>>> I'd leave it as shown in the attached diff fragment, except that I'm
>>> a bit worried about possible platform dependency of the output.  The
>>> hashing occurring in the subplans shouldn't affect output order, but
>>> I'm not sure if we want a test output like this or not.  Thoughts?
>
>> How about replacing "a = 3" with "a < 7 AND a != 6". That then
>> exercises more of the possible types of behaviour for quals: The "a <
>> 7" qual is pushed down and used as an index condition. The "a != 6"
>> qual is pushed down and used as a filter, because it's cheap and
>> leakproof. The leakproof() qual isn't pushed down on cost grounds. The
>> snoop() qual isn't pushed down on security grounds. Both snoop() and
>> leakproof() are used as filters, along with "a != 6", and a SB subplan
>> qual. "a != 6" is executed first because it has a security_level of 0,
>> and is cheaper than the subplan. snoop() is executed later, despite
>> being cheaper than the other filter quals, because it has a higher
>> security_level, and leakproof() is executed last because it has the
>> same security level as snoop() but is more expensive.
>
> Will do, although I think that the next test case (the one with "a = 8")
> already shows most of those behaviors.
>

Except that it doesn't have a cheap leakproof qual like "a != 6", not
handled automatically by the index, that order_qual_clauses() can
assign security_level = 0 to, and then move to the start of the list.

I think it's probably worth having a clause like that in one of the
tests, but it could perhaps be added to the "a = 8" test, if you
wanted to drop the "a = 3" test.

Regards,
Dean