Re: [PATCH 6/6] push: honor branch.*.push

On Mon, Jun 24, 2013 at 4:19 PM, Ramkumar Ramachandra <artagnon@gmail.com> wrote:

I haven't followed this topic closely, and the concern described below is probably a consequence of that. Please ignore if my worries are unfounded.

It seems to me like we're adding (or have already added) quite a bit of config variables and command-line options for specifying (at varying levels of specificity and overridability) either the remote to push to ($remote), or what ref to push into on the remote ($remote_ref). It worries me that we allow $remote and $remote_ref to be set _separately_ and at separate levels of specificity. I wonder if this too easily allows users to shoot themselves in the foot by specifying $remote in one place (e.g. on the command line), then (in their mind - unrelated) specifying $remote_ref in another place (e.g. branch.foo.push), and then being negatively surprised when the two conspire to push into the "wrong" $remote and/or $remote_ref.

I haven't yet dug deep enough to figure out an obvious failure mode (and I probably should not have sent this email until I'd found one), but I wonder if we'd be better off forcing the $remote and $remote_ref configuration for a given branch to appear as more of a single unit.

What if, when setting up tracking for a given branch, we immediately specified its complete pull and push targets?

For example, when in a centralized workflow (e.g. push.default = upstream) and we're checking out local branch foo from origin's foo, we could set up the following configuration [1]:

[branch "foo"]
    pull = origin/foo
    push = origin/foo

In a triangular workflow (assuming we had configuration to specify such, and also a default push remote), we could then instead set up the following config:

[branch "foo"]
    pull = origin/foo
    push = my_public/foo

This leaves no ambiguity for even the most novice user as to the pull and push targets for a given branch, and it's also easy to change it, either by editing the config file directly, or by using hypothetical commands:

  git branch foo --pulls-from=origin/bar
  git branch foo --pushes-to=other_repo/foo

Any "git pull" without arguments will fail if branch.<current>.pull is unset or invalid. Likewise any "git push" without arguments will fail if branch.<current>.push is unset or invalid.

Obviously, specifying the remote and/or refspec on the command-line would still override, as it does today, but for the argument-less forms of "git pull" and "git push", the hierarchy of options and defaults being consulted to figure out $remote and $remote_ref would be small and easily understandable.

...Johan

[1]: I do realize that I'm abusing the "foo/bar" notation to mean "$remote/$ref", and that this does not work in the general case where both remote names and ref names may contain slashes, or when remote names don't correspond to eponymous ref namespaces...