Re: nested loop semijoin estimates

I wrote:
> Hm.  In principle, add_path_precheck shouldn't be doing anything except
> rejecting paths that would get rejected anyway later on.  However it
> strikes me that there's a logic error in it; specifically, this
> simplifying assumption is wrong:
>          * For speed, we make exact rather than fuzzy cost comparisons. If an
>          * old path dominates the new path exactly on both costs, it will
>          * surely do so fuzzily.

After staring at this awhile longer, I still don't see exactly why it's
changing the results.  Do we have instructions around here anyplace on
how to set up/use TPC-DS?  I couldn't find anything about it on the
wiki ...
        regards, tom lane