On Wed, May 15, 2019 at 09:32:31AM -0400, Tom Lane wrote:
> David Fetter <david@fetter.org> writes:
> > I hope the patch is a little easier to digest as now attached.
>
> To be blunt, I find 500K worth of changes in the regression test
> outputs to be absolutely unacceptable, especially when said changes
> are basically worthless from a diagnostic standpoint.
You're right, of course. The fundamental problem is that our
regression tests depend on (small sets of) fixed strings. TAP is an
alternative, and could test the structure of the output rather than
what really should be completely inconsequential changes in its form.
> There are
> at least two reasons why this won't fly:
>
> * Such a change would be a serious obstacle to back-patching
> regression test cases that involve explain output.
>
> * Some buildfarm members use nonstandard settings (notably
> force_parallel_mode, but I don't think that's the only one).
> We are *not* going to maintain variant output files to try to cope
> with all those combinations. It'd be even more disastrous for
> private forks that might have their own affected settings.
Indeed. I think we should move our regression tests to TAP and
dispense with this.
> I don't know how to make progress towards the original goal without
> having a regression-test disaster, but what we have here is one.
This just highlights a disaster already in progress. I'm volunteering
to help fix it.
Best,
David.
--
David Fetter <david(at)fetter(dot)org> http://fetter.org/
Phone: +1 415 235 3778
Remember to vote!
Consider donating to Postgres: http://www.postgresql.org/about/donate