Re: [PATCH v1 2/2] worktree: make add dwim

[Thomas Gummerer <t.gummerer@gmail.com>]

On 11/14, Eric Sunshine wrote:
> On Tue, Nov 14, 2017 at 3:45 AM, Thomas Gummerer <t.gummerer@gmail.com> wrote:
> > On 11/13, Junio C Hamano wrote:
> >> If so, as long as the new DWIM kicks in ONLY when "topic" does not
> >> exist, I suspect that there is no backward compatibility worries.
> >> The command line
> >>
> >>     $ git worktree add ../a-new-worktree topic
> >>
> >> would just have failed because three was no 'topic' branch yet, no?
> >
> > Indeed, with this there would not be any backwards compatility
> > worries.
> >
> > Ideally I'd still like to make
> >
> >     $ git worktree add ../topic
> >
> > work as described above, to avoid having to type 'topic' twice (the
> > directory name matches the branch name for the vast majority of my
> > worktrees) but that should then come behind a flag/config option?
> > Following your reasoning above it should probably be called something
> > other than --guess.
> >
> > Maybe --guess-remote and worktree.guessRemote would do?  I'm quite bad
> > at naming though, so other suggestions are very welcome.
> 
> For my own edification...
> 
> git worktree add ../dir branch
> 
> * Checks out branch into ../dir if branch exists.
> 
> * Errors out if branch does not exist. However, if we consider the
> history of the implementation of worktrees[*1*], then this really
> ought to try the "origin/branch -> branch" DWIM before erroring-out.
> Implementing this DWIM could be considered a regression fix according
> to [*1*] and, as Junio points out, should not pose backward
> compatibility worries.

Agreed, I think it is not very controversial that this would be an
improvement.

> git worktree add ../topic
> 
> * Correctly errors out, refusing to create a new branch named "topic",
> if "topic" is already a branch.
> 
> * Creates a new branch named "topic" if no such local branch exists.
> 
> The desired new DWIMing would change the second bullet point to:
> 
> * If no local branch named "topic" exists, DWIM it from "origin/topic"
> if possible, else create a new local branch named "topic".
> 
> But that's a behavior change since, with the existing implementation,
> a newly created local "topic" has no relation to, and does not track,
> any origin/topic branch. The proposed --guess option is to avoid users
> being hit by this backward incompatibility, however, one could also
> view the proposed DWIMing as some sort of bug fix since, at least
> some, users would expect the DWIMing since it already takes place
> elsewhere.

I'm not sure we can call it a bug fix anymore, since as Junio pointed
out the current behaviour of creating a new branch at HEAD is
documented in the man page.

However git-worktree is also still marked as experimental in the man
page, so we could allow ourselves to be a little bit more lax when it
comes to backwards compatibility, especially because it is easy to
take corrective action after the new DWIMing happened.

> So, at least two options exist:
> 
> * Just make the new DWIMing the default behavior, accepting that it
> might bite a few people. Fallout can be mitigated somewhat by
> prominently announcing that the DWIMing took place, in which case the
> user can take corrective action (say, by choosing a different worktree
> name); nothing is lost and no damage done since this is happening only
> at worktree creation time.
> 
> * Add a new option to enable DWIMing; perhaps name it -t/--track,
> which is familiar terminology and still gives you a relatively short
> and sweet "git worktree add -t ../topic".
> 
> Given the mentioned mitigation factor and that some/many users likely
> would expect it to DWIM "origin/topic -> topic" anyhow, I'd lean in
> favor of the first option (but perhaps I'm too daring with other
> people's workflows).

Yeah, I'm leaning towards the first option as well, but I'm clearly
biased as that's how I'd like it to behave, and others might want the
other behaviour.  Unfortunately I don't know many worktree users, so I
can't tell what the general consensus on this would be.

I guess the second option would be the safer one, and we can still
switch that to be the default at some point if we wish to do so
later.

tl;dr I have no idea which of the options would be better :)

> FOOTNOTES
> 
> [*1*]: When Duy first implemented worktree support, he incorporated it
> directly into the git-checkout command ("git checkout --to worktree
> ..."), which means that he got all the git-checkout features for free,
> including the "origin/branch -> branch" DWIM. When worktree support
> was later moved to git-worktree, it lost most of the features
> inherited implicitly from git-checkout, such as -b, -B, --detach, so
> those were added back to git-worktree explicitly. However, at that
> early stage, git-worktree was still piggy-backing atop git-checkout,
> thus likely was still getting the "origin/branch -> branch" DWIM for
> free. A final iteration converted git-worktree away from heavyweight
> git-checkout to lightweight git-reset, at which point he DWIMing was
> lost. If you take this history into account, then loss of
> "origin/branch -> branch" DWIMing is a regression, so restoring it
> could be considered a bug fix.