Re: What's cooking in git.git (Dec 2015, #01; Tue, 1)

[Patrick Steinhardt <ps@pks.im>]

[-- Attachment #1: Type: text/plain, Size: 4586 bytes --]

On Mon, Dec 07, 2015 at 11:24:52AM -0800, Junio C Hamano wrote:
> Patrick Steinhardt <ps@pks.im> writes:
> 
> > On Wed, Dec 02, 2015 at 05:31:14PM -0500, Jeff King wrote:
> >> On Wed, Dec 02, 2015 at 02:11:32PM -0800, Junio C Hamano wrote:
> > [snip]
> >> > "--keep-empty" has always been about keeping an originally empty
> >> > commit, not a commit that becomes empty because of rebasing
> >> > (i.e. what has already been applied to the updated base).  The
> >> > documentation, if it leads to any other interpretation, needs to be
> >> > fixed.
> >> > 
> >> > Besides, if "--keep-empty" were to mean "keep redundant ones that
> >> > are already in the updated base", the patch must do a lot more,
> >> > e.g. stop filtering with git-cherry patch equivalence.
> >> > 
> >> > I'm inclined to eject this topic.
> >> 
> >> That was my thinking too (and I notice it didn't get any review from
> >> anybody else).
> > [snip]
> >
> > Well, I kind of agree. The cherry-pick command has both
> > --allow-empty and --keep-redundant flags, where the second one is
> > the kind of behavior I want to achieve in my case. As an
> > alternative to the proposed change to `--keep-empty` I could
> > instead introduce a new flag `--keep-redundant-commits` to `git
> > rebase` which would then pass the flag through to the
> > cherry-pick.
> >
> > Any opinions on this?
> 
> Honestly, I am not interested if that is only about passing that
> option and doing nothing else.
> 
> Imagine that you have two changes from the branch being rebased
> already in the updated base, one was accepted verbatim, while the
> other one was accepted with a slight tweak.  Perhaps there were some
> conflicts in the context, or something.
> 
> When you run your rebase, the former will not even be considered
> because command will notice, via patch equivalence, that you do not
> need it, even before it attempts to replay it.  But not the latter.
> The command will attempt to replay it, and only in this step it may
> notice, by seeing that the result becomes a no-op change, that the
> change has already been included in the updated base.
> 
> I cannot quite see how adding "--keep-redundant-commits" to the
> command would help anybody by allowing the latter in the history but
> not the former.  That is primarily because you can imagine a future
> in which the patch equivalence determination becomes smarter.  With
> or without "--keep-redundant-commits", both of these two changes
> would not appear in the resulting history when that happens.
> 
> The "--keep-redundant" option to "cherry-pick" makes sense, as the
> user will be the one who is deciding which commit to replay on top
> of a new base.  If the user fed a commit that is already in the new
> base, and the command is run with the option, that is a deliberate
> request to leave a no-op cruft.
> 
> We cannot say the same thing for "rebase", as it tries to avoid
> redundant cruft, and it cannot do as good a job as humans in doing
> so.  The "--keep-redundant-commits" option will become a way to make
> a distinction between what got dropped by the command automatically
> and what didn't get dropped because the command did not look deeply
> enough.  That distinction is not at all useful, as that can change
> as the tool improves.
> 
> A "--keep-redundant-commits" to "rebase" that also disables the
> patch equivalence filtering (not just keeping empty crufts in the
> resulting history) might make sense, but I do not see myself (or
> anybody sane) using it.  "In what situation would such a feature be
> useful?" is a question I cannot answer offhand.

The scenario I require this feature for is somewhat more exotic,
yes. We've got a CI where published branches can be rebased but
should not be modified in their commit history. That is, the CI
determines in a hook wether the branch that is being pushed drops
or re-orders commits and if so, rejects the branch.

So when a commit that is already included in origin/master gets
rebased `git-rebase` currently simply drops the commit, which
causes the CI to refuse the branch on a push.

So obviously you are right in your conclusion that
`--keep-redundant-commits` has to skip the patch equivalence
determination in order to not drop any commits. Otherwise my
change would only work for certain scenarios.

That said, I do not care too deeply about this patch, especially
as my scenario is rather uncommon. If you deem this to not have
any greater benefit you may simply drop it.

Patrick

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 819 bytes --]