On Fri, Apr 29, 2016 at 7:33 PM, Tom Lane <
tgl@sss.pgh.pa.us> wrote:
>
> Amit Kapila <
amit.kapila16@gmail.com> writes:
> >> On Thu, Apr 28, 2016 at 10:06 PM, Tom Lane <
tgl@sss.pgh.pa.us> wrote:
> >>> I'd be inclined to think that it's silly to build GatherPaths in advance
> >>> of having finalized the list of partial paths for a rel.
>
> > What's happening here is that to form joinrel, we need to call
> > add_paths_to_joinrel() with both outer and inner relation twice, once with
> > rel1 as outer relation and rel1 as inner relation and vice versa. So now
> > the second call to add_paths_to_joinrel() can replace a partial path which
> > is being referenced by GatherPath generated in first call. I think we
> > should generate gather paths for join rel after both the calls
> > to add_paths_to_joinrel() aka in make_join_rel(). Attached patch on above
> > lines fixes the problem for me.
>
> make_join_rel is certainly not far enough down the call stack to solve
> this problem. It can, and typically will, be invoked multiple times
> for the same target join relation.
>
> One possible answer is to do it in standard_join_search, just before
> the set_cheapest call for each join relation.
>
Yes, that makes sense to me.
> You'd need to account
> for the issue in GEQO search as well.
>
How about doing it in merge_clump() before calling set_cheapest()?
Yet, another idea could be to create a copy of partial path before passing it to create_gather_path()?