Re: [RFC PATCH] Introduce "precious" file concept

From: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Date: Nov 28, 2018, 21:54 UTC
Message-ID: <875zwgzx4v.fsf@evledraar.gmail.com>
In-Reply-To: <xmqqa7ltua2s.fsf@gitster-ct.c.googlers.com>

On Wed, Nov 28 2018, Junio C Hamano wrote:

Show 10 quoted lines

> Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:
>
>> What do you think about some patch like that which retains the plumbing
>> behavior for things like read-tree, doesn't introduce "precious" or
>> "trashable", and just makes you specify "[checkout|merge|...] --force"
>> in cases where we'd have clobbering?
>
> Whether you like it or not, don't people's automation use tons of
> invocations of "git merge", "git checkout", etc.?  You'd be breaking
> them by such a change.

I'm so sympathetic to this argument that I tried to convince you of something like this around a year and a half ago: https://public-inbox.org/git/CACBZZX59KXPOEjiUKtZLN6zjO_xpiWve7Xga6q-53J2LwvfZyw@mail.gmail.com/ :)

I was probing for what your current stance on this sort of thing is, because discussions like this tend to get bogged down in the irrelevant distraction of whether something is plumbing or porcelain, which almost none of our users care about, and we've effectively stopped caring about ourselves.

But we must have some viable way to repair warts in the tools, and losing user data is a *big* wart.

I don't think something like the endgame you've described in https://public-inbox.org/git/xmqqzhtwuhpc.fsf@gitster-ct.c.googlers.com/ is ever going to work. Novice git users (the vast majority) are not going to diligently update both .gitignore and some .gitattribute mechanism in lockstep. I'd bet most git users haven't read more than a few paragraphs of our entire documentation at best.

So what's the way forward? I think ultimately we must move to something where we effectively version the entire CLI UI similar to stable API versions. I.e. for things like this that would break some things (or Duy's new "split checkout") introduce them as flags first, then bundle up all such flags and cut a major release "Git 3, 4, ...", and eventually remove old functionality.

Show 26 quoted lines

> Other than that, if we never had Git before and do not have to worry
> about existing users, I'd think it would be a lot closer to the ideal
> than today's system if "checkout <tree> foo.o" rejected overwriting
> "foo.o" that is not tracked in the current index but matches an ignore
> pattern, and required a "--force" option to overwrite it.
>
> A user, during a conflict resolution, may say "I want this 'git
> checkout foo/' to ignore conflicted paths in that directory, so I
> would give "--force" option to it, but now "--force" also implies
> that I am willing to clobber ignored paths, which means I cannot use
> it".
>
> I would think that a proper automation needs per-path hint from the
> user and/or the project, not just a single-size-fits-all --force
> option, and "unlike all the *.o ignored files that are expendable,
> this vendor-supplied-object.o is not" is one way to give such a
> per-path hint.
>
>> This would give scripts which relied on our stable plumbing consistent
>> behavior, while helping users who're using our main porcelain not to
>> lose data. I could then add a --force option to the likes of read-tree
>> (on by default), so you could get porcelain-like behavior with
>> --no-force.
>
> At that low level, I suspect that a single size fits all "--force"
> would work even less well.

Yeah I don't think the one-size-fits-all way out of this is a single --force flag.