On 06/19/2018 10:55 AM, Tom Lane wrote:
AFAICS, the fast-default patch neglected to consider what happens if
a database containing columns with active attmissingval entries is
pg_upgraded. I do not see any code in either pg_dump or pg_upgrade that
attempts to deal with that situation, which means the effect will be
that the "missing" values will silently revert to nulls: they're still
null in the table storage, and the restored pg_attribute entries won't
have anything saying it should be different.
The pg_upgrade regression test fails to exercise such a case. There is
only one table in the ending state of the regression database that has
any atthasmissing columns, and it's empty :-(. If I add a table in
which there actually are active attmissingval entries, say according
to the attached patch, I get a failure in the pg_upgrade test.
This is certainly a stop-ship issue, and in fact it's bad enough
that I think we may need to pull the feature for v11. Designing
binary-upgrade support for this seems like a rather large task
to be starting post-beta1. Nor do I think it's okay to wait for
v12 to make it work; what if we have to force an initdb later in
beta, or recommend use of pg_upgrade for some manual catalog fix
after release?
Ouch!
I guess I have to say mea culpa.
My initial thought was that as a fallback we should disable pg_upgrade
on databases containing such values, and document the limitation in the
docs and the release notes. The workaround would be to force a table
rewrite which would clear them if necessary.
Have we ever recommended use of pg_upgrade for some manual catalog fix
after release? I don't recall doing so. Certainly it hasn't been common.
I have no idea how large an actual fix might be. I'll at least start
working on it immediately. I agree it's very late in the day.
cheers
andrew
