Re: gc and repack ignore .git/*HEAD when checking reachability

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi Duy,

On Wed, 13 Jul 2016, Duy Nguyen wrote:

> On Wed, Jul 13, 2016 at 10:20 AM, Johannes Schindelin
> <Johannes.Schindelin@gmx.de> wrote:
> >
> > On Tue, 12 Jul 2016, Duy Nguyen wrote:
> >
> >> On Tue, Jul 12, 2016 at 5:51 PM, Jeff King <peff@peff.net> wrote:
> >> > On Tue, Jul 12, 2016 at 05:46:12PM +0200, Duy Nguyen wrote:
> >> >
> >> >> I'm not opposed to letting one worktree see everything, but this move
> >> >> makes it harder to write new scripts (or new builtin commands, even)
> >> >> that works with both single and multiple worktrees because you refer
> >> >> to one ref (in current worktree perspective) differently. If we kill
> >> >> of the main worktree (i.e. git init always creates a linked worktree)
> >> >> then it's less of a problem, but still a nuisance to write
> >> >> refs/worktree/$CURRENT/<something> everywhere.
> >> >
> >> > True. I gave a suggestion for the reading side, but the writing side
> >> > would still remain tedious.
> >> >
> >> > I wonder if, in a worktree, we could simply convert requests to read or
> >> > write names that do not begin with "refs/" as "refs/worktree/$CURRENT/"?
> >> > That makes it a read/write-time alias conversion, but the actual storage
> >> > is just vanilla (so the ref storage doesn't need to care, and
> >> > reachability just works).
> >>
> >> A conversion like that is already happening, but it works at
> >> git_path() level instead and maps anything outside refs/ to
> >> worktrees/$CURRENT.
> >
> > Wouldn't you agree that the entire discussion goes into a direction that
> > reveals that it might simply be a better idea to require commands that want
> > to have per-worktree refs to do that explicitly?
> 
> No.

Oh? So you do not see that we are already heading into serious trouble
ourselves?

> I prefer we have a single interface for dealing with _any_ worktree.

I agree. So far, I did not see an interface, though, but the idea that we
should somehow fake things so that there does not *have* to be an
interface.

> > The same holds true for the config, BTW. I really have no love for the
> > idea to make the config per-worktree. It just holds too many nasty
> > opportunities for violate the Law of Least Surprises.
> >
> > Just to name one: imagine you check out a different branch in worktree A,
> > then switch worktree B to the branch that A had, and all of a sudden you
> > may end up with a different upstream!
> 
> Everything in moderation. You wouldn't want to enable sparse checkout
> on one worktree and it suddenly affects all worktrees because
> core.sparsecheckout is shared. And submodules are known not to work
> when core.worktree is still shared.

We have precendence for config variables that are global but also allow
per-<something> overrides. Think e.g. the http.* variables.

The point is: this solution still uses *one* config per repo.

> I will not enforce any rule (unless it's very obvious that the other
> way is wrong, like core.worktree). I will give you a rifle and you can
> either hunt for food or shoot your foot. In other words, you should be
> able to share everything if you like it that way while someone else
> can share just some config vars, or even nothing in config file.

You gave me a rifle alright, and I shot into my foot (by losing objects to
auto-gc). I just did not expect it to be a rifle.

To keep with the analogy: let's not build arms, but a kick-ass tool. And I
seriously disagree that per-worktree refs, reflogs or config are part of
said tool.

Ciao,
Dscho