hubert depesz lubaczewski <depesz@depesz.com> writes:
> Worked a bit to get the ltree problem down to smallest possible, repeatable, situation.
I looked at this again and verified that indeed, commit
8eee65c996048848c20f6637c1d12b319a4ce244 introduced an incompatible
change into the on-disk format of ltree columns: it widened
ltree_level.len, which is one component of an ltree on disk.
So the crash is hardly surprising. I think that the only thing
pg_upgrade could do about it is refuse to upgrade when ltree columns
are present in an 8.3 database. I'm not sure though how you'd identify
contrib/ltree versus some random user-defined type named ltree.
regards, tom lane