On Sat, Apr 26, 2025 at 8:53 PM Peter Geoghegan <pg@bowt.ie> wrote:
On Sat, Apr 26, 2025 at 10:39 PM Mark Dilger <mark.dilger@enterprisedb.com> wrote: > Peter, Matthias, thanks kindly for the good work on skipscans!
Thanks!
> I found a test case which fails after commit 21a152b37f36c9563d1b0b058fe1436baf578b4c. Please find a reproducible test case, attached.
The bug isn't actually in commit 21a152b37f36c9563d1b0b058fe1436baf578b4c -- it's just an accident that the mechanism added by that commit happens to make your test case fail. The underlying issue was introduced in commit 8a510275, "Further optimize nbtree search scan key comparisons".
This looks to have been a silly oversight in our handling of NULL tuple datums within _bt_check_compare. Attached provisional fix makes your test case pass.
-- Peter Geoghegan
I can confirm that your patch fixes the problem, having spent some four hours trying to find other test cases which still fail, finding none.
Thank you for the quick reply, and again for the work on btree.
