Re: [HACKERS] Runtime Partition Pruning

On 11 April 2018 at 09:32, Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
> Robert Haas wrote:
>> On Mon, Apr 9, 2018 at 2:28 PM, Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
>
>> >> I don't get this. The executor surely had to (and did) open all of
>> >> the relations somewhere even before this patch.
>
>> > I was worried that this coding could be seen as breaking modularity, or
>> > trying to do excessive work. However, after looking closer at it, it
>> > doesn't really look like it's the case. So, nevermind.
>>
>> Well what I'm saying is that it shouldn't be necessary. If the
>> relations are being opened already and the pointers to the relcache
>> entries are being saved someplace, you shouldn't need to re-open them
>> elsewhere to get pointers to the relcache entries.
>
> I looked a bit more into this. It turns out that we have indeed opened
> the relation before -- first in parserOpenTable (for addRangeTableEntry),
> then in expandRTE, then in QueryRewrite, then in subquery_planner, then
> in get_relation_info.
>
> So, frankly, since each module thinks it's okay to open it every once in
> a while, I'm not sure we should be terribly stressed about doing it once
> more for partition pruning. Particularly since communicating the
> pointer seems to be quite troublesome.

I guess the problem there would be there's nothing to say that parse
analysis will shortly be followed by a call to the planner, and a call
to the planner does not mean the plan is about to be executed. So I
don't think it would be possible to keep pointers to relcache entries
between these modules, and it would be hard to determine whose
responsibility it would be to call relation_close().

It might be possible to do something better in each module by keeping
an array indexed by RTI which have each entry NULL initially then on
first relation_open set the element in the array to that pointer.

This might mean we'd save a few relation_open calls, but I don't know
if there would be a way to somehow remove the Relation from the array
on relation_close. Having something like this might mean we could
detect lock upgrade hazards more easily, but the whole thing is a
cache on top of a cache which does seem a bit weird. relation_open()
should be pretty cheap if the relation is already open. It's just a
hash table lookup. What is described above just changes that to an
array lookup. It also does nothing for index_open.

However, something like the above would simplify
ExecLockNonLeafAppendTables() a bit and get rid of the O(N^2) which
checks the partition is not a result relation.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services