On 08/12/2025 10:41, Michael Paquier wrote:
> On Sun, Dec 07, 2025 at 11:57:34PM -0500, Tom Lane wrote:
>> Nah, hold off till we have a discussion about whether the test is
>> correct to expect compatibility. We can afford to have a couple
>> of BF animals unhappy for a day or two while we come to a decision.
>
> By the way, while looking at the buildfarm failures, I have noticed
> that something is wrong in my patch. Looking Makefile.global.in,
> USE_MODULE_DB is built based on REGRESS for the tests of
> src/test/modules/gin/. However things are inconsistent between meson
> and ./configure.
>
> For example, this host uses configure/make:
> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=skimmer&dt=2025-12-08%2006%3A08%3A23
>
> The database name is contrib_regression_gin_incomplete_splits, which
> points to the test gin_incomplete_split in src/test/modules/gin/,
> meaning that my patch would not work.
>
> In the meson case, we got this failure:
> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=crake&dt=2025-12-08%2006%3A27%3A04
> And here, the database name is regression_gin.
>
> When I looked at the buildfarm failures, I saw the regression_gin
> first, wrote a patch based on that. But the second one based on
> contrib_regression_gin_incomplete_splits was really puzzling me, until
> I've noticed that the difference was configure vs meson for the two
> hosts. So it seems to me that we have a second issue here, where
> meson relies on a "name" while configure uses the first value of
> "REGRESS". AdjustUpgrade.pm cannot know that it's dealing with one
> build method or the other. In order to make my previous patch work,
> we could use a DROP DATABASE IF EXISTS on both database names. That
> would work, but that feels grotty for the purpose of fixing things,
> especially as I don't see a point in keeping the test libraries across
> major upgrades..
Hmm, using the first value of REGRESS is surprising. How about we change
the autoconf logic to match what meson does? As attached.
In the meson scripts, you can specify the database name explicitly by
setting 'dbname'. It's only used in a few subdirs though.
This in src/test/interfaces/ecpg/test/meson.build looks weird:
> 'dbname': 'ecpg1_regression,ecpg2_regression',
As far as I can tell, dbname is not a list, so I don't understand how
that works.
- Heikki
