The point I'm making is that the scenario your test case exposes is not an infinite loop of picksplits, but an infinite loop of choose calls.
Thank you, now I see what I missed before. After some brainstorming, it's possible to use '\0' only as end of string marker. The idea is: when we split allthesame tuple to upper/lower we copy node label to upper tuple instead of set it to '\0'. Node labels of lower tuple will be unchanged but should be ignored for reconstructionValue - and they are ignored because of allTheSame flag. See attached patch.
Unfortunately, it will break on-disk compatibility. Although it will not cause a panic/error/coredump allTheSame approach will not find tuples. Users should recreate spgist indexes over text column.
It's certainly possible that we could/should change what checkAllTheSame is doing --- on re-reading the comment, I'm not really sure that the scenario it's concerned about can happen. However, that will not fix
I rewrited a patch to fix missed way - allTheSame result of picksplit and tooBig is set. I believe this patch is still needed because it could make a tree more efficient as it was demonstrated for quad tree.
this patch doesn't break on-disk compatibility, although index recreation is recommended.
--
Teodor Sigaev E-mail: teo...@sigaev.ru
WWW: http://www.sigaev.ru/
spgist_textproc.patch.gz
Description: Unix tar archive
spgist_allthesame.patch.2.gz
Description: Unix tar archive
-- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
