Alexander Korotkov <aekorotkov@gmail.com> writes:
> On Thu, Feb 16, 2012 at 11:43 AM, Alexander Korotkov
> <aekorotkov@gmail.com>wrote:
>> Described differences leads to incorrect behaviour of GiST index.
>> The question is: what is correct way to fix it? Should on_pb also use FP*
>> or consistent method should behave like on_pb?
> Any comments on this? Current behaviour definitely indicates a bug, and I'm
> ready to fix it. The only question: is this bug in on_pb or gist?
I'm inclined to think the right answer is to make on_pb use the FP*
macros, for consistency with other geometric operators. But it's worth
asking whether that will actually fix the problem. I've thought for
some time that we'd eventually find cases where geo_ops' use of fuzzy
comparisons breaks index behavior entirely, because it destroys natural
assumptions like the transitive law. So that could eventually lead us
to rip out the FP* macros everywhere.
In any case, this doesn't seem like something we could back-patch;
it'd be a behavioral change in HEAD only.
regards, tom lane