Michael Paquier <michael@paquier.xyz> writes:
> On Mon, Sep 03, 2018 at 06:59:10PM -0700, Noah Misch wrote:
>> If you're going to keep this highly-simplified estimate, please expand the
>> comment to say why it doesn't matter or what makes it hard to do better. The
>> non-planunionor.c path for the same query computes its own estimate of the
>> same underlying quantity. Though it may be too difficult in today's system,
>> one could copy the competing path's row count estimate here. Perhaps it
>> doesn't matter because higher-level processing already assumes equal row count
>> among competitors?
> As there has been no replies to Noah's review for one month, I am
> marking this patch as returned with feedback for now.
FWIW, my problem with this patch is that I remain unconvinced of the basic
correctness of the transform (specifically the unique-ification approach).
Noah's points would be important to address if we were moving the patch
towards commit, but I don't see much reason to put effort into it until
we can think of a way to prove whether that works.
regards, tom lane