Re: pg_upgrade test for binary compatibility of core data types

On Wed, Dec 01, 2021 at 04:19:44PM +0900, Michael Paquier wrote:
> I'll get that done down to 10 to maximize its influence, then I'll
> move on with the buildfarm code and send a patch to plug this and
> reduce the dependencies between core and the buildfarm code.

Okay, I have checked this one this morning, and applied the split down
to 10, so as we have a way to fix objects from the main regression
test suite. The buildfarm client gets a bit cleaned up after that (I
have a patch for that, but I am not 100% sure that it is right).

Still, the global picture is larger than that because there is still
nothing done for contrib/ modules included in cross-version checks of
pg_upgrade by the buildfarm. The core code tests don't do this much,
but if we were to do the same things as the buildfarm, then we would
need to run installcheck-world (roughly) on a deployed instance, then
pg_upgrade it. That's not going to be cheap, for sure.

One thing that we could do is to use unique names for the databases of
the contrib/ modules when running an installcheck, so as these are
preserved for upgrades (the buildfarm client does that). This has as
effect to increase the number of databases for an instance
installcheck'ed, so this had better be optional, at least.
--
Michael