Robert Haas <robertmhaas@gmail.com> writes:
> On Tue, Jul 24, 2018 at 7:51 AM, Etsuro Fujita
> <fujita.etsuro@lab.ntt.co.jp> wrote:
>> I'm not sure that's a good idea, because I think we have a trade-off
>> relation; the more we make create_plan simple, the more we need to make
>> earlier states of the planner complicated.
>>
>> And it looks to me like the partitionwise join code is making earlier (and
>> later) stages of the planner too complicated, to make create_plan simple.
> I think that create_plan is *supposed* to be simple. Its purpose is
> to prune away data that's only needed during planning and add things
> that can be computed at the last minute which are needed at execution
> time. Making it do anything else is, in my opinion, not good.
I tend to agree with Robert on this. In particular, if you put anything
into create_plan or later that affects cost estimates, you're really doing
it wrong, because changing those estimates might've changed the plan
entirely. (As I recall, we have a couple of cheats like that now,
associated with dropping no-op projection and subquery scan nodes.
But let's not add more.)
TBH, I think this entire discussion is proving what sheer folly it was
to allow partitions to have rowtypes not identical to their parent.
But I suppose we're stuck with that now.
regards, tom lane