On Sun, Jul 25, 2010 at 6:40 PM, Greg Stark <gsstark@mit.edu> wrote:
> 2010/7/25 Robert Haas <robertmhaas@gmail.com>:
>> 2010/7/25 PostgreSQL - Hans-Jürgen Schönig <postgres@cybertec.at>:
>>>
>>> On Jul 25, 2010, at 11:56 AM, Martijn van Oosterhout wrote:
>>>
>>>> I think the right way to approach this is to teach the planner about
>>>> merge sorts.
>
> For what it's worth I think this is a belt-and-suspenders type of
> situation where we want two solutions which overlap somewhat.
>
> I would really like to have merge-append nodes because there are all
> sorts of plans where append nodes destroying the ordering of their
> inputs eliminates a lot of good plans. Those cases can be UNION ALL
> nodes, or partitions where there's no filter on the partition key at
> all.
>
> But for partitioned tables like the OPs the "real" solution would be
> to have more structured meta-data about the partitions that allows the
> planner to avoid needing the merge at all. It would also means the
> planner wouldn't need to look at every node; it could do a binary
> search or equivalent for the right partitions.
Agreed on all points.
>> Greg Stark had a patch to do this a while back called merge append,
>> but it never got finished...
>
> I was basically in over my head with the planner. I don't understand
> how equivalent classes are used or should be used and didn't
> understand the code I was pointed at as being analogous. It's probably
> not so complicated as all that, but I never really wrapped my head
> around it and moved onto tasks I could make more progress on.
Yeah, I don't fully understand those either.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company