On Wed, Oct 29, 2025 at 6:43 AM Arne Roland <[email protected]> wrote: > Richard already covered a lot. I mainly want to reiterate, that a public test > case would be immensely helpful.
I agree, and said the same. > The Q1 you mentioned sadly isn't a real test case, where I can measure > performance impact. More an academic difference in costs, which I don't fully > comprehend as of now. I don't either, but maybe if I study it (or you or someone else does) we can begin to comprehend it. > Did you encounter a case a in production, that made you reevaluate this > thread? If so a public reproducer would be very appreciated. No, what happened is that this broke a patch I'm working on. The details are lengthy and would take us too far away from the topic of this thread, but the highly-compressed version is that I spent about six hours going "wait, why the heck isn't this working?" and eventually traced it back to the pathlist for a partitionwise join getting zapped. I might have to bolt on some kind of a fix to un-break that patch for now, but it's not relevant in terms of constructing a reproducer for this problem. I think the best shot at coming up with a reproducer here is to study the cost differences in the queries where the plan changes with the fix, particularly Q1 from my prior email. While I agree with you that at present that is just a numerical effect and not a real performance effect, we don't even have an explanation for how the numerical effect is possible. It seems like a good idea to figure that out. -- Robert Haas EDB: http://www.enterprisedb.com
