Re: [PATCH] BreakingChanges: early adopter option

From: Junio C Hamano <gitster@pobox.com>
Date: Sep 20, 2024, 21:33 UTC
Message-ID: <xmqqploydn7j.fsf@gitster.g>
In-Reply-To: <xmqq7cb77810.fsf@gitster.g>

Junio C Hamano <gitster@pobox.com> writes:

Show 17 quoted lines

> Discussing the desire to make breaking changes, declaring that
> breaking changes are made at a certain version boundary, and
> recording these decisions in this document, are necessary but not
> sufficient.  We need to make sure that we can implement, test, and
> deploy such impactful changes.
>
> Formalize the mechanism based on the `feature.*` configuration
> variable to allow early adopters to opt into the breaking change in
> a version of Git before the planned version for the breaking change.
>
> Signed-off-by: Junio C Hamano <gitster@pobox.com>
> ---
>
>  * Before I forget.  I'll find time to rewrite the "we no longer
>    honor core.preferSymlinkRefs" topic to follow this new guideline
>    when we see a rough concensus that both the procedure outlined
>    here and the idea to remove core.preferSymlinkRefs are good.

So I was looking at my other topic that prepares to deprecate and remove core.preferSymlinkRefs configuration (i.e. we would still be able to work with a repository whose HEAD is a symbolic link pointing at refs/heads/main, but we would never be able to use a symbolic link to represent a symref ourselves, so "git checkout next" would replace HEAD with a textual symref, a file that has a single line "ref: refs/heads/next" in it), to see what it involves to allow early adopters to experience Git 3.0 features/removals before it actually happens.

Switching behaviour at runtime with feature.git3 should work well, and we can also add tests that checks the new behaviour by doing "test_config feature.git3 true".

One trouble is what to do with "git $cmd --help" for those who choose to be early adopters.

For "git $cmd --help" and "git -c feature.git3=true $cmd --help" to show documents describing the behaviour before and after Git 3.0 version boundary, we'd need to install separate set of documents and switch between them at runtime in builtin/help.c:setup_man_path() and friends [*]. If we are going to have such a transition often (e.g., more frequently than every few years), laying such an infrastructure would certainly be worth it, but an obvious alternative is to say that, even though the toggling of behaviour can be done at runtime to allow the early adopters from going back to the current behaviour to make their life less risky, the contents of the manual pages are defined at the compile time.

If we are willing to burden early adopters a bit more, we could make it a build-time option. With "make GIT_BUILD_FOR_GIT3=YesPlease", binaries will be built for all the then-current Git 3.0 features and documentation under development. It certainly is a simpler-to-build option that is easier for us, but I am not sure if that is acceptable by those who volunteer to test the upcoming big version.

One thing to note is that depending on the nature of a change, once you start using a feature only available in a newer version of Git in your repository, the resulting repository may not be understood by an older version of Git (imagine you started using index v4---now you are locked out of your repository if you want to go back to a version of Git before 1.7.11). So in that sense, it might be a disservice to promise that an early adopter can experience the future with feature.git3=on and then can safely go back to the current world by flipping it off, and we might be better off to control this with a single big red build-time switch.

Opinions?

[Footnote]

 * Alternatively we could tweak system_path() to hack the value we
   return when asked about GIT_INFO_PATH, GIT_MAN_PATH, and
   GIT_HTML_PATH, but I somehow feel dirty just for mentioning such
   an implementation possibility X-<.