configuring cherry-pick to always use -x?

Is there a configuration option to make cherry-pick always include the source commit hash in the new commit log message?

e.g., make "git cherry-pick" always behave like "git cherry-pick -x"?

My most frequent use case for cherry picking is between publicly visible branches.

I have the following configuration option set:

alias.cpx=cherry-pick -x

but I rarely remember to use it.

Not currently, but before we go any further, could you please justify in what workflow it would make sense to use -x most of the time?

We used to add the "cherry-picked from" by default in the very early days, and stopped doing so for a reason, and also deliberately stayed away from adding such a configuration to actively discourage the use of -x without making the user thinking twice.

Sure. Summary: two long-lived publicly visible branches.

Details:
Mifos is what I'm usually working on lately. We have branches "master" and "f-release" both present in our public git repository called "head" (hosted at sf.net). master is the bleeding edge of development, f-release is a release maintenance branch recently created off the tip of master. I expect both to live on forever (even though commits to f-release will eventually cease).

Right after f-release was cut, we merged f-release to master every day or so to make sure bugfixes for f-release were also propagated to future releases. After a while, merging resulted in too many conflicts and we started cherry picking instead.

This process is described generally at
http://mifosforge.jira.com/wiki/display/MIFOS/Release+Branch+Merging+Policy
.

If the source commit is present in the log message of the new (cherry picked) commit, it's easy to (1) find the source commit (gitweb creates a hyperlink, for instance) and (2) know that, when viewing the log of master, a particular commit is also present on another branch. Right now I just keep reminding folks to use -x.

For (2), I generally assume that branch is a release branch, but come to think of it, it would be nice to know what branch a commit was cherry picked from. For example: "(cherry picked from BRANCHNAME commit c6e08938e352f3ec99a29a67dd192945d2bcf00d)" would be better than the current message generated by -x.

See also:
http://mifosforge.jira.com/wiki/display/MIFOS/Mifos+Version+Control+Guide

Adam Monsen venit, vidit, dixit 14.02.2011 22:50:

I don't quite understand how cherry picks could conflict less then merges if the release branch contains fixes only. Also, I don't think the advice to use "merge+revert" is a good one. All of this indicates a suboptimal use of branches. My impression is that "f-release" actually mixes release engineering and maintenance. Two possible remedies:

- Separate release engineering from maintenance and merge only the latter to master

- If you do want them on the same branch "f-release", you probably know beforehand which commits you don't want on master. You can fake-merge these ("merge -Xours") to master and merge the others, which is somewhat ugly but still better than cherry-picking everything. In some sense this is "manual rerere" whose results are shared (pushed) easily.(*)

Michael

(*) If that is cryptic, I mean something like:

git checkout master
git merge f-release
#be happy if it succeeds, identify problematic commit X if not; decide whether X belongs on master; if yes resolve, if not reset and:
git merge X^
git merge -Xours X
#back to start

For the record, I think that should be -sours.

I think it's just a typo but the difference is big --- -sours means "supersede by pretending to merge but actually keeping our version", while -Xours means "do a normal merge but be sloppy and favor our change when encountering adjacent or overlapping changes".

I suppose -Xours should have been named -Xfavor-ours, -Xsloppy-favoring-ours, or something similarly explicit.

Thanks for a nice example.
Jonathan

Jonathan Nieder venit, vidit, dixit 15.02.2011 10:18:

Yes, sorry and thanks.

-Xours may still be useful to know for the OP, but for the complete fake-merge you need -sours. In fact, "-sours" has an awfully good mnemonic when used for fake-merging commits which you do not want to cherry-pick :)

Michael

The last time I experienced a painful merge from f-release to master, it was because some files had been culled from master but left extant on f-release. Not too hard to resolve, actually. But I really only needed one change pulled into master, and when I cherry picked instead of merging the whole branch, there were no conflicts, and master ended up containing exactly what I wanted.

Ah, thank you! This is invaluable advice. I think I'll go with this option since mixing release engineering and maintenance is exactly what I'm doing. Hopefully it's worth the added complexity of having another public branch.

I pushed an example to https://github.com/meonkeys/releaseBranchDemo
that I'll share with my developers.

"git merge -sours" will definitely be something useful to add to the quiver too.

I personally prefer cherry-pick, as the fake merge clutters mainline's history with the superseded commits.

It would be interesting to have a history simplification that given a merge M of parents A and B, ignored commits $(merge-base A B)..B where M is TREESAME to A. Hmm, that's effectively "git rev-list ." isn't it?

j.

In one of my repos, most of the time my cherry-picks are between two public branches. Perhaps a better enhancement would be something like:

branch.<name>.annotate_cherry_pick = {true, false}

which could be set to true for source branches that you wish to default to -x. Or, maybe it makes sense in cases where the source branch is a remote-tracking branch:

cherry_pick.annotate = {local, remote}

I'm not sure how good a remote-tracking branch is as an indicator of 'public branch', though, so I think explicitly configuring it per-branch makes more sense. I hesitate there only because we don't currently put remote-tracking branches in the branch section names.

j.

Hi Junio,

Could you elaborate on the reason why -x is a bad idea?

Kind regards,

Ivan Kanis, Release Manager, Vision Objects,
Tel +33 2 28 01 84 44,  Fax +33 2 40 25 89 20
http://www.visionobjects.com

We meet no Stranger, but Ourself.
    -- Emily Dickinson 

Nope, but one would be appreciated. :-)

It's worse than that. I like to keep the message generated after a cherry-pick conflict, but the original commit authorship. I have this, which I call recommit:

<snip>
#!/bin/sh
# Used after a cherry-pick conflicts to commit with the original

# authorship (commit -c) but keep the newly generated commit message
#
self=$(cd "$(dirname "$0")" && pwd -P)/$(basename "$0")
. "$(git --exec-path)/git-sh-setup"
require_work_tree
cd_to_toplevel
test -f .git/MERGE_MSG || die "No .git/MERGE_MSG"

if test "$GIT_EDITOR" = "$self"
then
cat .git/MERGE_MSG > .GIT/COMMIT_EDITMSG
exit 0
fi

if sha1=$(sed -ne \
's/^(cherry picked from commit \([a-f0-9]\{40\}\))$/\1/p' .git/MERGE_MSG)
then
export GIT_EDITOR="$self"
git commit -c $sha1
fi
</snip>

I've had it on my TODO list for a while now to:

1. add a config option to enable -x by default

2. improve the cherry-pick conflict UX. I was thinking of out CHERRY_HEAD on conflict and then adding a cherry-pick --continue option which acts like rebase --continue. CHERRY_HEAD is what was being picked at the time of conflict and can be used by the bash completion script for proper prompting, as well as obviously the --continue option.

j.

Yes, we have so far only "use commit -c $that_one", which is an obvious and low hanging fruit for improvement. Thanks.