At Tue, 19 May 2015 09:04:00 -0400, Robert Haas <robertmh...@gmail.com> wrote in <CA+TgmobAV3_DS1sXA+PFWkjvX1K-VNiAnMMJrzPfD43g=-4...@mail.gmail.com> > On Tue, May 19, 2015 at 7:19 AM, Andrew Gierth > <and...@tao11.riddles.org.uk> wrote: > >>>>>> "Tom" == Tom Lane <t...@sss.pgh.pa.us> writes: > > Tom> Hm. That's a hangover from when query_planner also gave back a > > Tom> Plan (singular) rather than a set of Paths. I don't see any > > Tom> fundamental reason why we couldn't generalize it to be a list of > > Tom> potentially useful output orderings rather than just one. But I'm > > Tom> a bit concerned about the ensuing growth in planning time; is it > > Tom> really all that useful? > > > > The planning time growth is a possible concern, yes. The potential gain > > is eliminating one sort step, in the case when the input has a usable > > sorted path but grouping_planner happens not to ask for it (when there's > > more than just a single rollup, the code currently asks for one of the > > sort orders pretty much arbitrarily since it has no real way to know > > otherwise). Whether that would justify it... I don't know. Maybe that's > > one to save for later to see if there's any feedback from actual use. > > I kind of doubt that the growth in planning time would be anything too > unreasonable. We already consider multiple orderings for ordinary > base relations, so it's not very obvious why consideration multiple > orderings for subqueries would be any worse. If we can arrange to > throw away useless orderings early, as we do in other cases, then any > extra paths we consider have a reasonable chance of being useful.
Though I don't think that the simple path-ification of what is currently done make it grow in any degree, it could rapidly grow if we unconditionally construct extra upper-paths using the previously-abandoned extra paths or make them involved in join considerations. But the growth in planning time could be kept reasonable if we pay attention so that, as we have done so, the additional optimization schemes to have simple and precise bailing-out logic even if they require some complicated calculation or yields more extra paths. regards, -- Kyotaro Horiguchi NTT Open Source Software Center -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
