Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
> I'm about to do some benchmarking on -HEAD one some hardware I have
> available and it seems I'm hitting a rather weird issue causing the osdl
> dbt3 benchmark to run very slow and eating CPU time for hours ...
Could we see the actual EXPLAIN ANALYZE results for the slow plan?
I'm unconvinced by your "hash join is bad" analysis, especially in
the cases where you're giving it lots of work_mem. I think it's got
something to do with the different join orders. The rowcount estimates
in the fast plan all seem pretty good, but I'm betting something is
wrong with some of them in the slow case.
regards, tom lane