Re: Rangejoin rebased

Hi Jeff,

Just a quick comment -- I ran a slightly modified version of a query
from the regression tests, and got an assertion failure:

select i1, ir1, i2, ir2
  from (select * from rangejoin_left order by ir1 desc) as a1 inner
join (select * from rangejoin_right order by ir2 desc) as a2
    on (i1 = i2 and ir1 && ir2)
  order by ir1 desc, i1;
TRAP: FailedAssertion("!(!ssup->ssup_reverse)", File:
"/home/akuzmenkov/pgp-old/build/../postgrespro/src/backend/executor/nodeMergejoin.c",
Line: 492)

The sort order isn't right for the join, it seems. I remember having
similar troubles with my full merge join implementation. I tried
filtering unsuitable paths in try_mergejoin_path, but that was not quite
enough. The planner tries to use only one sort direction to limit the
number of path, so the path we need might not be there at all. This
optimization was added in commit 834ddc62, see right_merge_direction().
Sadly, I have no idea how to solve this.

--
Alexander Kuzmenkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company