Re: `git fetch` not updating 'origin/HEAD' after branch rename

From: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Date: Apr 14, 2021, 10:56 UTC
Message-ID: <87sg3t16ec.fsf@evledraar.gmail.com>
In-Reply-To: <CAPx1GvdeNEyPEZ7GdRKeAevnvjyLmoXHjQP0W6iToDsJPAqDHA@mail.gmail.com>

On Wed, Apr 14 2021, Chris Torek wrote:

Show 21 quoted lines

> On Tue, Apr 13, 2021 at 5:18 PM Sam Bostock <sam.bostock@shopify.com> wrote:
>> Long story short, it seems to me that `git fetch` should update
>> "refs/remotes/origin/HEAD" when the upstream HEAD changes, but it
>> doesn't.
>
> It's never been the *intent* to have `git fetch` update
> the corresponding remote-tracking `HEAD` ref.  To make
> that happen, you must run `git remote`:
>
>     git remote set-head origin -a
>
> for instance.
>
> I have, however, often thought that this is the wrong
> *default* way for things to work, and that at least by default,
> `git fetch origin` should update `origin/HEAD` if the
> fetch result indicates that it should.  See also Junio's
> reply.  I think a configuration knob (similar to `fetch.prune`)
> would be reasonable here.  Users could then be encouraged
> to set `fetch.prune` to `true`, and `fetch.update-remote-HEAD`
> (or whatever) to `true` as well.

As you'll see from the previous linked thread I happen to agree with you, but let's not step past that agreement and (no doubt subtly and unintentionally, in this case) misrepresent the other side, which also has a legitimate argument.

Which is that this notion that "[a client should update its view of a remote to what a] fetch result indicates that it should" isn't something that exists, and that we ever had it is only an emergent implementation detail.

So first of all, a "fetch result" doesn't look anything like that. Rather the server presents refs it has, and the client decides what it wants.

Which is the start of the disconnect in mental models around this. What you see with "git ls-remote origin" != what you should expect to have at a local refs/remotes/origin/* after a fetch.

Nothing in git itself actually needs this remote HEAD past clone time, and with 4f37d457065 (clone: respect remote unborn HEAD, 2021-02-05) there'll be even less reason to pay attention to it.

We don't even always get it on "clone", we don't do it with "--bare", since the reason we do it is to one-time setup the default branch, which we then hardcode in the config in the non-bare case:

    [remote "origin"]
            url = https://github.com/git/git.git
            fetch = +refs/heads/*:refs/remotes/origin/*
    [branch "master"]
            remote = origin
            merge = refs/heads/master

So arguably the bug is the other way around, that we should never save this information at all in the ref store at "clone" time, which would make "init && config && fetch" and "clone" consistent.

I haven't looked, but wouldn't be surprised to discover that this was originally needed back when we needed to ferry information between shellscripts, and was just diligently retained when the relevant parts were rewritten in C.

But I digress.

Now, what people *do actually* legitimately use this information is a convenient local cache of "what's the main upstream branch?".

I myself have written local scripts that used that, and been bitten by this information being stale / not there (in the case of init/config/fetch).

But why does anyone need that? Well, I don't think they actually do. What they actually do want is to push or fetch the "main" branch. The "what was HEAD that one time I talked to this remote" is just a roundabout way of getting that.

So isn't this whole thing just wart that we should fix in the protocol? Wouldn't the use-case for this be satisfied with something like:

    [branch "master"]
            remote = origin
            merge = HEAD ; does not work as you might expect

Right now that will just push to refs/heads/HEAD, but what if we had a protocol extension to intercept it (or rather, some merge = <a name incompatible with a current push, maybe "$HEAD">, as an aside setting it to ":HEAD" has some very funny results) wouldn't that satisfy the use-case?

After all, who's really interested in what the remote's idea of their HEAD when they last fetched is?

Don't those users actually want the *current* idea of what HEAD is for the purposes of fetching or pushing?

If we supported "I want the tip of your HEAD" (which we do) or "I have this ref update for you, on top of your idea of HEAD" (we don't) you could "git push" and it would do the right thing whether the primary branch was renamed from "master" to "main" during your "git push" or not.