Andrew Dunstan <andrew.dunstan@2ndquadrant.com> writes:
> Thanks, Committed. Further investigation shows this was introduced in
> release 12, so that's how far back I went.
Still further investigation shows that this patch caused bug #16655 [1].
It should *not* have been designed to unconditionally clear the
table's "interesting" flag, as there may have been other reasons
why that was set. The right way to think about it is "if we are
going to dump the table's data, then the table certainly needs its
interesting flag set, so that we'll collect the per-attribute info.
Otherwise leave well enough alone".
The patches I proposed in the other thread seem like they really ought
to go all the way back for safety's sake. However, I do not observe
any crash on the test case in v11, and I'm kind of wondering why not.
Did you identify exactly where this was "introduced in release 12"?
regards, tom lane
[1] https://www.postgresql.org/message-id/flat/16655-5c92d6b3a9438137%40postgresql.org