Nathan Bossart <nathandbossart@gmail.com> writes:
> On Fri, Aug 04, 2023 at 09:28:51AM +1200, David Rowley wrote:
>> Thank you for reviewing. I've pushed the patch to master only.
> I'm seeing some reliable test failures for 32-bit builds on cfbot [0]. At
> a glance, it looks like the relations are swapped in the plan.
Yeah, I got the same result in a 32-bit FreeBSD VM. Probably, the two
plans are of effectively-identical estimated cost, and there's some
roundoff effect in those estimates that differs between machines with
4-byte and 8-byte MAXALIGN.
You could likely stabilize the plan choice by joining two tables that
aren't of identical size -- maybe add an additional WHERE constraint
on one of the tables?
regards, tom lane