Re: [PATCH 1/3] branch: introduce --set-upstream-to

Jonathan Nieder [off-list ref] writes:

[someone should have]
| suggested an alternative syntax that avoids the mistake you quoted
| above, perhaps something like:
|
| 	git branch --set-upstream-to=origin/master [HEAD]

with which I disagree.

You can think of it this way.

"git branch" can not only _create_ a new branch (or list existing
ones, but that is another entirely different mode), but also can be
used to set attributes to an existing branch.  Imagine a new option,
say --set-description, to replace branch.frotz.description, for
example.  It would be used like this:

	$ git branch --set-description='add frotz feature' frotz

to set the description for the 'frotz' branch (i.e. the above would
set branch.frotz.description), and we default to HEAD if 'frotz' is
missing from the command line.  "git branch --option [<branch>]" is
about manipulating the branch, and we default the target of
manipulation to HEAD.

"upstream" is just another kind of attribute for the branch being
manipulated, whose value happens to be a branch name.

The mistake was that --set-upstream was coded by piggybacking the
existing --track implementation where a new branch was created, and
in that codepath, "git branch <name1> [<name2>]" creates <name1>
while defaulting a missing <name2> to HEAD.

Creating a new branch that is forked from the current HEAD is an
often useful thing to do, so defaulting a missing <name2> (aka
"start-point") to HEAD is very sensible, but reconfiguring a named
branch <name1> to integrate with the current branch is much less
useful than the other way around.  One major reason why it is so is
because you would more likely set any branch to integrate with a
remote tracking branch (rather than a local branch) and by
definition your HEAD cannot be a remote tracking branch.

It makes it worse that you would often want to reconfigure the
current branch; for the purpose of reconfiguring a branch <name1> to
integrate with something else <name2>, it is much more likely that
you want a missing <name1> to default to HEAD, not the other way
around to default a missing <name2> to HEAD, which is useful for
branch creation.

But switching which missing argument gets default based on what
options are used is insane.

If the very original "create this new branch starting at that point"
were spelled like this

	$ git branch [--start-point=<name2>] <name1>

and a missing <name2> defaulted to HEAD, it probably would have been
better. It would have made it very unlikely to tempt anybody to hack
the --set-upstream option into the system with the wrong parameter
order if such a command line convention was in place.

If anything, it could be a sensible longer-term direction to a more
intuitive UI to deprecate the two-name format and make the creation
to be specified with an explicit --start-point option with an
argument (which defaults to HEAD), but I think that falls into the
"if I were reinventing git without existing userbase in 2005"
category and it is too late for that.

Junio C Hamano wrote:

You can think of it this way.

"git branch" can not only _create_ a new branch (or list existing
ones, but that is another entirely different mode), but also can be
used to set attributes to an existing branch.  Imagine a new option,
say --set-description, to replace branch.frotz.description, for
example.  It would be used like this:

	$ git branch --set-description='add frotz feature' frotz

That's the same question.

You say that it would be used like that.  I say that it would be
more intuitive, given how "git remote", "git config", and other
commands other than "update-index --chmod" that set attributes already
work, for it to be used like this:

	git branch --set-description frotz 'add frotz feature'

Notice how similar that is to "git remote set-head origin master".
It would just be the consistent thing to do.

The truth is that neither one of us is right.  Both conventions
could work, and which one is more intuitive will vary from person
to person.  The convention used for plain "git branch" is

	copy(target, source)

That matches memcpy() and is the opposite of what "cp" uses.  Oh
well.  The convention used for "git remote add" is

	method(this, args...)

It's generally pretty natural.  The convention used for "git
update-index --chmod" is

	action(parameters)(files...)

That matches "chmod" so it was probably a good choice.

Hoping that clarifies,
Jonathan