On 24/05/2023 10:33, Tao Klerks wrote:
On Wed, May 24, 2023 at 2:06 AM Junio C Hamano <gitster@xxxxxxxxx> wrote:
Tao Klerks <tao@xxxxxxxxxx> writes:
The current implementation of this patch is far too restrictive. It
doesn't break any tests (and maybe I should add one now that I know),
but it's doing the wrong thing.
I am ambivalent. What do we want to see in a multi-pick sequence
that is different from rebase?
I would argue there are primarily three things that are different:
1. The checkout of the new base (and checkout of the original in an "--abort")
2. The support for and/or more-common expectation of "messing" with
commits as you go, eg squash, edit
3. The (partial) support for rebasing/recreating merge commits
I'm not sure to what extent any of these justify having tighter
restrictions on when we allow a rebase to start, though.
A single-step cherry-pick can fail
safely before it touches the index or the working tree files, but if
two-step cherry-pick, whose first step succeeds, finds that it
cannot safely carry out its second step without clobbering the local
changes made to the working tree files, what should happen? Are we
OK if we stopped in the state just after the first step has already
been done?
This is the current behavior: it stops before the specific pick that
is going to affect local unstaged changes, or if there are *any*
staged changes (in which case it stops as it's about to do the first
pick - the first time this check runs). The reasoning for this
behavior, as I understand it, is that the "--abort" strategy,
intending to "undo whatever I started doing here, including a conflict
resolution", resets the index. So as long as there is nothing you want
to keep in the index, and as long as we know that any previous picks
haven't impacted any files with unstaged changes, we're good.
"cherry-pick --abort" uses "reset --merge" so is slightly more
complicated than just resetting the index.
The bug that I want to fix is that we only end up checking whether
there are changes in the index *after* we've already committed to
resetting the index upon later "--abort". It's a kind of catch-22:
we've detected that aborting would destroy your work, so we leave you
in a state where the most obvious thing to do is abort, so we destroy
your work... Of course, if you understand what's going on you can
choose to "--quit" and *not* lose your work... but this is completely
antithetical to the general intent of "--abort".
There's another, smaller flaw here I think, common to Merge,
Single-Cherry-Pick, and Sequence-Cherry-Pick, which is that *if* you
start with unstaged changes, and you end up in a conflict resolution
or "--no-commit" pause, and you then "git add" your unstaged changes
during that pause/resolution, and you *then* later "--abort"... then
your originally-unstaged changes are destroyed by the "--abort" - so
it has *not* taken you back to where you were before the operation
started. This is, to me as a user, non-obvious, and could potentially
lead to data loss. The only way I see to fix that, is to have *all* of
these operations refuse to operate on dirty worktrees altogether -
like rebase already does.
I suspect this level of "strictness" would be welcome to newcomers,
and less welcome to existing experienced users.
Indeed, being able to cherry-pick when the worktree is dirty is
occasionally useful but it is confusing when "cherry-pick --abort" does
not restore the pre-cherry-pick state (and as Junio says below it would
be complicated to preserve the initial changes when aborting).
My (tentative) answer to that question is "yes", but the recovery
options of "cherry-pick" may want to work differently from what we
have seen them traditionally do. Namely, the user accepts that the
first step is already done, and stopping "cherry-pick", be it called
"--abort" or something else, should just remove the sequencer state
and behave as if the single-pick cherry-pick on the first step only
has just finished and leave such a state in the index and the
working tree.
This behavior exists, and is called "--quit", right?
I think so though I'm not sure what what "--quit" does to CHERRY_PICK_HEAD.
The semantics as I understand it are:
--quit: I know what I'm doing, just remove any "ongoing operation"
metadata and let me work with the current index and worktree.
--abort: This was a bad idea, please take me back to where I was
before I started this operation (without losing any work I had
ongoing, pls!)
If that is what we are going to do, then it would
make sense to adopt the same safety semantics we use for "git merge"
and "git checkout" to ensure only that the index is clean, relying
on the unpack-trees machinery that stops before clobbering a locally
modified working tree files.
Yep
But if we are to aim for "all-or-none"
semantics people expect from aborting "git rebase", I suspect that
it would be way too complicated to allow random changes in the
working tree files that we may only discover to be problems after
starting the sequence of replaying commits one-by-one, and "too
restrictive" check may be justified.
>
I don't think I understand this argument. If we want to support both
sets of semantics, then that's exactly what "--quit" and "--abort"
achieve, right? (as long as we check for the dirty index *before*
committing to destroying the index in case of "--abort")
To put it differently, if it
is too restrictive for multi-pick, then we would want to loosen it
for "git rebase" as well, as the issues are likely to be the same.
My argument for only changing "sequence-cherry-pick" here, and having
it (continue to) use the index-safety-only semantics of
single-cherry-pick and merge, is that *this is not a change in
cabability* - it is only a bugfix. Switching to the worktree-safety
semantics of rebase would be a substantial change in behavior beyond
the bugfix.
I, personally, would prefer to see the worktree-safety semantics of
rebase be used in *all* these operations, so I could no longer shoot
myself in the foot by starting a merge, accidentally staging some
previously-unstaged changes during conflict resolution, and then
losing those changes by "--abort"ing. But I expect that this kind of
change would need to be behind a config option of some sort, trading
off safety against low friction.
I could imagine a setting like "core.OperationWorktreeSafety", with
settings "default" (current behavior - rebase disallows dirty
worktrees, the others disallow dirty index), "strict" (all behave like
current rebase) and "lax" (all behave like merge).
As discussed elsewhere, I would also like to (have an option to) treat
untracked files as "worktree dirtiness"/unstaged changes in exactly
the same way as changes to tracked files - but that's another topic :)
I'll prepare a v2 with index-safety-only for sequence-cherry-pick for
now, please let me know if a (better-named)
"core.OperationWorktreeSafety" option is something that you'd be
interested in / that would make sense to you.
I think it's worth considering a config option. One problem though is
that if the user has to actively enable it to get greater safety it wont
be much help to new users who don't know about it and if it is on by
default we'll be breaking someone's existing workflow.
Best Wishes
Phillip
Thanks!
