Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:
> This only happens on some runs (in my case the CREATE INDEX passed 8x
> without any issue and failed on 9th iteration), so there has to be some
> random element affecting this.
As far as that goes, gistchoose() intentionally makes random choices
when two insertion paths seem equally good. So failing only sometimes
isn't very surprising.
> so maybe the GIST stack is stale, not updated, or something like that. I
> think I've seen an issue mentioning that recently ... yep, found it [1].
> And I see Tom came to about the same conclusion, pointing to the same
> place in gistfinishsplit. It's only a couple days old, I don't think
> I've seen any proposed fixes yet.
Good that we seem to have just one bug there not two :-(. I've not
looked any further than my previous message, though.
I wonder if this could be a recently-introduced bug? I do not
recall seeing complaints like this before v12.
regards, tom lane