Re: pg15b2: large objects lost on upgrade

On Thu, Aug 4, 2022 at 12:59 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
> Why you think it's better to not have the test than to have a very limited
> amount of fuzziness (by using the next xid as an upper limit). What's the bug
> that will reliably pass the nextxid fuzzy comparison, but not an exact
> comparison?

I don't know. I mean, I guess one possibility is that the nextXid
value could be wrong too, because I doubt we have any separate test
that would catch that. But more generally, I don't have a lot of
confidence in fuzzy tests. It's too easy for things to look like
they're working when they really aren't.

Let's say that the value in the old cluster was 100 and the nextXid in
the new cluster is 1000. Well, it's not like every value between 100
and 1000 is OK. The overwhelming majority of those values could never
be produced, neither from the old cluster nor from any subsequent
vacuum. Given that the old cluster is suffering no new write
transactions, there's probably exactly two values that are legal: one
being the value from the old cluster, which we know, and the other
being whatever a vacuum of that table would produce, which we don't
know, although we do know that it's somewhere in that range. Let's
flip the question on its head: why should some hypothetical future bug
that we have in this area produce a value outside that range?

If it's a failure to set the value at all, or if it generates a value
at random, we'd likely still catch it. And those are pretty likely, so
maybe the value of such a test is not zero. On the other hand, subtle
breakage might be more likely to survive developer testing than
complete breakage.

--
Robert Haas
EDB: http://www.enterprisedb.com