Robert Haas <robertmhaas@gmail.com> writes:
> On Sun, Jul 3, 2016 at 1:14 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I think a cleaner way is to have set_append_rel_size() invoke
>> set_rel_consider_parallel() on the child rels and then propagate their
>> parallel-unsafety up to the parent. That seems fairly analogous to
>> the way we already deal with their sizes: in particular, set_rel_size
>> is invoked on an appendrel child from there, not from an extra pass in
>> set_base_rel_sizes. Then set_append_rel_pathlist can propagate unsafety
>> down again, as you've done here.
> I was reluctant to do it that way because of the relatively stupid way
> that we have to go about finding the inheritance children to visit,
> but maybe I shouldn't let that dominate my thinking. We could always
> replace that with some more efficient data structure at some point
> down the road.
That might be a reasonable objection if we needed to add a third such
loop, but I'm talking about putting the set_rel_consider_parallel call in
the loop that already exists. So there shouldn't be any added overhead.
(It's definitely true that we could improve on the append_rel_list-based
search mechanism; but so far I have never seen any indication that that
was a bottleneck, so it's pretty far down my to-do list.)
>> You mentioned that you'll be on vacation for much of July. If you like,
>> I will take this open item off your hands, since I'll be around and can
>> deal with any bugs that pop up in it.
> That would be much appreciated.
OK, will do.
regards, tom lane