Re: [PATCH v2 0/1] Make 'git commit' not accidentally lose staged content

From: Jeff King <peff@peff.net>
Date: Sep 17, 2018, 18:15 UTC
Message-ID: <20180917181511.GA3894@sigill.intra.peff.net>
In-Reply-To: <CACsJy8C5QOLvg4pzy_pThQoyGh9ohdeVHXsuYwQHQypn3oBxkw@mail.gmail.com>

On Mon, Sep 17, 2018 at 07:29:26PM +0200, Duy Nguyen wrote:

Show 25 quoted lines

> > On the other hand, if I am keeping some change that should never be
> > in a commit in the working tree file, and building the contents in
> > the index using "add -p" to incrementally, it would be the same
> > disaster as you are trying to prevent if I by mistake did a whole
> > path 'add', even if I catch myself doing so before running 'commit'
> > i.e.
> >
> >         edit X
> >         git add -p X
> >         git diff --cached X
> >         git diff X
> >         ... repeat the above number of times ...
> >         git add X ;# OOPS!
> >         git add . ;# OOPS! even worse!
> >
> > Even though this does not involve "git commit -a" or "git commit X",
> > an unrecoverable damage that requires redoing the manual work is
> > already done.
> 
> I don't see a good way to get to recover this situation. I could go
> back to the "index log" idea, where we keep a log of index changes (or
> just "interesting" changes). That way there's no behavior change at
> all. The user who accidentally updates/deletes something can always
> retrieve the old content back (assuming that they realize quickly
> since we can't keep very long log).

FWIW, I like that approach much better, since:

  1. It does not bother or restrict anybody in their workflow; instead,
     they pay the complexity price only when they know they have made a
     mistake.

  2. It covers many more cases (e.g., just doing the wrong thing via
     "add -p").

A naive index log would be pretty cheap in CPU, at least for POSIX-ish systems. You could just hard link "index" to "index.N" before renaming "index.lock" over "index". But I guess if you have a gigantic index, that's less appealing. So maybe storing the equivalent of a "--raw" diff between the two index states would make more sense (and after all, you don't really need the stat-cache or cache-tree). It would cost more to reconstruct the index on the fly, but then the point is that you would create these logs a lot more than you access them.

> I've been thinking about allowing to undo worktree changes too (e.g.
> accidental "git reset --hard") and this log can cover it as well.

I like that, too. It's a little more costly just because it may involve object-db writes, but I think in most cases it would be fine. I almost always "git stash" away discarded changes these days instead of "git reset --hard", because it effectively provides this kind of log.

> The only downside is we need a new command for the UI (or perhaps I
> can just add "git add --log" or something like that).

I think the reflog approach has been successful: give these intermediate states a name. So in theory I could do something like:

  git checkout -p :@{2.minutes.ago}

though it would probably be useful to be able to walk the states, too, just like we have "log --reflog-walk".

The syntax above is off-the-cuff (and based on the ":<stage>" index syntax). I guess if we had a separate log for "stuff in the worktree that got thrown away by reset" we might need a separate syntax.

> Should I just drop this patch and go that direction instead? More
> work, but maybe better end result.

I guess my whole response is a long-winded "yes". ;)

-Peff