On Wed, Mar 22, 2017 at 03:26:21PM -0700, Junio C Hamano wrote:
> Ævar Arnfjörð Bjarmason <avarab@xxxxxxxxx> writes:
>
> > of things you think we should be putting in the test suite. I.e.
> > should the tests be:
> >
> > a) Only be a collection of invocations of git we'd be comfortable
> > showing to someone as "this works, and this is how you should do it",
> > or things that explicitly fail marked with test_must_fail.
> >
> > b) or a) && also various surprising combinations of things we don't
> > necessarily want to encourage or even support in the future, but which
> > are in there so if we change them, we at least know our change changed
> > something that worked before.
>
> I am strongly inclined to (a). If we cannot decide when we designed
> the feature, and we anticipate that we may want to change it later,
> then documenting the choice in a test or two may be a way to remind
> the choice we happened to have made, but in general I do not think
> we want to promise (to ourselves) more than what we are willing to
> commit to.
I've occasionally[1] added tests that are "what we happen to produce
now", but I almost always mark them with a comment either in the test
script or in the commit message. What I'm _most_ concerned about is a
developer later breaking the test, but being unsure if they were
breaking some real-world case (and not being able to find clues in the
history).
A secondary concern would be people using the test snippets as guidance
on what is normal or encouraged.
So I could live with these patches, but I'd prefer to see a comment
somewhere. And I think I'd have a slight inclination to just stick to
(a) in the first place, unless there is a really good reason to cover
the test (like that we do not care between behaviors X and Y, but we
need to check that it does one of them, and not Z).
-Peff
[1] E.g., see the comment in t3204 from a356e8e2a (t3204: test
git-branch @-expansion corner cases, 2017-03-02).