Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> Basing an MV on pg_class could always be difficult for pg_upgrade. Maybe
> that's not a brilliant thing to do in a test (or maybe the test should drop
> the MV after it's done).
OH. Is that what it's doing? The cause of the failure is immediately
obvious then. pg_class now lacks a relhaspkey column, but the matview
is still dependent on one being there.
I concur that this is not well-considered test design.
regards, tom lane