Web lists-archives.com

Re: [PATCH 0/2] read-tree: improve untracked file support




On 02/05/2019 11:53, Duy Nguyen wrote:
On Wed, May 1, 2019 at 9:58 PM Phillip Wood <phillip.wood123@xxxxxxxxx> wrote:
On 01/05/2019 11:31, Duy Nguyen wrote:
On Wed, May 1, 2019 at 5:14 PM Phillip Wood <phillip.wood123@xxxxxxxxx> wrote:

From: Phillip Wood <phillip.wood@xxxxxxxxxxxxx>

These two patches teach read-tree how to avoid overwriting untracked
files when doing '--reset -u' and also how to respect all of git's
standard excludes files. I'd like to see the porcelain commands stop
overwriting untracked files, this is a first step on the way. I'm not
sure if we want to add options to the porcelain commands to protect
untracked files or just change their behavior and add an option to
override that. I'm leaning towards the latter but I'd be interested to
hear what others think.

For new commands like git-restore, it's definitely a good thing to not
overwrite untracked files.

I agree, unfortunately this series does not help with git-restore, only
git-switch. For restore on an index without conflicts I think it could
just use the pathspec in struct unpack_trees_options and set opts.rest =
UNPACK_RESET_PROTECT_UNTRACKED but that does not help if we want to
handle conflicted paths differently to non-conflicted paths.

Right. I got confused. You did mention "git checkout <rev> :/" in 1/2,
which is the same as "git restore --source <rev> --staged --worktree
:/" and  can also potentially overwrite untracked files, even though
it does not use unpack-trees and cannot be fixed with this. Never
mind. Let's leave git-restore out of the discussion for now.

For existing commands I guess we have to go
over them one by one. For "git reset --hard", it should really just
overwrite whatever needed to get back to the known good state. "git
checkout -f" , not so sure (seems weird that we need force-level-two
option to override the protection provided by -f, if we change default
behavior)

I think it's fine for "checkout -f" to overwrite untracked files (and if
"switch --discard-changes" does not then there is no pressing need to
add such a mode to checkout), --force is a good name for an option that
nukes everything that gets in it's way. For "reset --hard" I'm not so
sure, if I have changes to an untracked file I don't wont them
overwritten by default. There is no porcelain equivalent to "read-tree
--reset --protect-untracked -u" and I was hoping "reset --hard" may
become that porcelain equivalent with a new --force or
--overwrite-untracked option.

My (biased, obviously) view is that "git reset --hard" is very
dangerous and I'm not trying to change that, especially when its
behavior has been like this since forever and I'm sure it's used in
scripts.

Instead "git restore" should be used when you need "git reset --hard
HEAD", the most often use case. And since it's new, changing default
behavior is not a problem. Which brings us back to git-restore :)

Does restore clean up the branch state like reset? It's tricky because you only want to do that if there is no pathspec (or the pathspec is :/ or equivalent - I can't remember if restore always requires paths or not)

But either way, git-restore or git-reset, I still don't see why
untracked files are more valuable in this case than tracked ones to
change the default.

My issue is that is easy to see what changes you're going to lose in tracked files by running diff. For untracked files diff just says a new file will be created, it ignores the current contents as the path is in the index so it is easy to overwrite changes without realizing. There's also a philosophical point that git should not be stomping on paths that it is not tracking though that's a bit moot if a path is tracked in one revision but not another.

I can see that sometimes you may want to restore
just tracked files, or untracked files, almost like filtering with
pathspec.

For the various "foo --abort" some (most?) are using "reset --merge"
which I think declines to overwrite untracked files but rebase uses
"reset --hard" which I'd like to change to protect untracked files in
the same way that rebase does for the initial checkout and when picking
commits. I haven't thought about stash.

Yeah it looks like cherry-pick and revert use "reset --merge" too
(reset_for_rollback function). That's all of them. Probably a stupid
question, why can't rebase just use "rebase --merge" like everybody
else?

I'm not sure - if --merge works for the others I can't see why it shouldn't work for rebase as well.

Best Wishes

Phillip
Best Wishes

Phillip