Re: "git checkout" safety feature

[Duy Nguyen <pclouds@gmail.com>]

On Mon, Nov 5, 2018 at 7:53 AM Jeff King <peff@peff.net> wrote:
>
> On Mon, Nov 05, 2018 at 07:24:42AM +0100, Matthias Urlichs wrote:
>
> > Hi,
> > > "git checkout <commit> <pathspec>" is a feature to overwrite local
> > > changes.  It is what you use when you make a mess editing the files
> > > and want to go back to a known state.  Why should that feature be
> > > destroyed?
> >
> > Not destroyed, but optionally made finger-fumble-save – like "alias rm
> > rm -i".
>
> There are a couple of destructive commands left in Git (e.g., this one,
> and "git reset --hard" is another). I didn't dig up archive references,
> but the topic of safety valves has come up many times over the years.
> The discussion usually ends with the notion that instead of warning
> that the operation is destructive (because that gets annoying when its
> purpose is to be destructive), we should make it possible to undo a
> mistake.
>
> So in this case, that would mean saving the working tree file to a blob
> before we obliterate it.
>
> See similar discussion in:
>
>   https://public-inbox.org/git/CACsJy8C5QOLvg4pzy_pThQoyGh9ohdeVHXsuYwQHQypn3oBxkw@mail.gmail.com/
>
> for example.

That work is still ongoing (slowly). I realized that reflog code was
buried deep in files-backend.c and would not make sense to reuse in
its current form. So I had to move the code to a common place, which
adds more work. But it will be coming! Hopefully before 2020 at my
usual development speed.

While we're at it, I've been running something with that "index
reflog" (no pruning) for a month with lots of "add -p" and the file
size is just 163KB, so the reflog format seems promising for this
purpose.
-- 
Duy