Re: [GENERAL] pg_upgrade from 9.5 to 9.6 fails with "invalid argument"

Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> On 09/29/2016 05:48 PM, Tom Lane wrote:
>> We might've caught these things earlier if the buildfarm testing
>> included cross-version upgrades, but of course that requires a
>> lot of test infrastructure that's not there ...

> I have done quite a bit of work on this - see
> <https://github.com/PGBuildFarm/client-code/blob/master/PGBuild/Modules/TestUpgradeXversion.pm>

Oh! How would I enable or use that?

> This actually runs on crake, and it has found problems in the past which
> we've fixed. However, it requires quite a deal of disk space (4GB or
> so), and the results are not stable, which is why I haven't enabled it.

Fair enough, but right now I'd like to do a one-shot test using a
big-endian machine to see whether it catches the apparent endianness
problem in rewriteVisibilityMap. I have a feeling that that is something
that won't easily be caught by automated checks, because it requires a
case where the table pages are in a variety of visibility states, and even
if corruption has happened, only index-only-scan queries would notice.

regards, tom lane