Tomas Vondra <tv@fuzzy.cz> writes:
> I've managed to further simplify the test-case, and I've verified that
> it's reproducible on current 9.2 and 9.3 branches.
Yeah, I've been able to reproduce it as well. I found what I thought
might be the bug: gbt_var_bin_union() looks like it does the wrong thing
if both ends of the range need to be extended. However, fixing that
doesn't fix your test case, so there's more to it.
> I've been unable to reproduce the bug with just a single text column.
Hm. I noticed a lot of elog complaints about "picksplit method
... doesn't support secondary split", so maybe the bug is in the code
that's trying to deal with that. I've not had time to look closer.
regards, tom lane