Re: cherry picking and merge

From: Jakub Narębski <jnareb@gmail.com>
Date: Aug 6, 2014, 15:43 UTC
Message-ID: <53E24D07.9010105@gmail.com>
In-Reply-To: <40F24BA38E03454A9BA152F6AFDE56C4@PhilipOakley>

Philip Oakley wrote:

> From: "Mike Stump" <mikestump@comcast.net>
> Sent: Friday, August 01, 2014 11:24 PM
>> On Aug 1, 2014, at 12:01 PM, Jakub Narębski <jnareb@gmail.com> wrote:

Show 12 quoted lines

>>> It can work in Subversion because Subversion stores information about
>>> what was merged in (and this includes cherry-picks, or whatever it is
>>> named in svn) in svn:mergeinfo property. Git does not track what was
>>> merged in, instead it represent the history as the graph of revisions,
>>> and tracks merges (by storing that it came from two or more commits)
>>> and not merged-in information.
>>
>> So, as a dumb user that just wants it to work, I am unsympathetic to
>> the `but software is hard’ excuse.  I am aware that some bugs are
>> harder to fix than others.  svn took a long time to fix this bug, but
>> they did.  I can wait, the only question is, will it be a week, a
>> month, a year, or a decade.

Here Git and Subversion went in different directions, and use different mechanisms (merge tracking vs merged-on tracking). Both have their advantages and disadvantages.

git-merge (in the most usual case) depends only on three revisions: the revision you merge into (current branch, ours), the revision you are merging (merged branch, theirs), and merge base (common ancestor). We could have another merge strategy that examines contents of revisions to handle cherry-picks and reverts... but it would be more complicated, and much slower.

Show 5 quoted lines

>>> When merging Git uses only what is being merged and its common
>>> ancestor (3-point merge). It is simple, and simple works!!!
>>
>> I gave a solution for git using branches and it works just fine.  It
>> retains the simple 3-point merge as well.

It works for this simple case, but I think it has unfortunate potential to go silently wrong.

Also, it prevents fully removing (commits, not only refs) the branch you cherry-picked from. The commit you cherry picked may no longer be (or may no longer should be) in the repository.

Show 15 quoted lines

> At the moment there is no formal way for Git to record within the commit
> metadata the inclusion of the cherry-picked diff (the 'merge' of the fix).
>
> Thinking out of the box, the issue is that the commit parents list does
> not have a formal mechanism to allow the recording that the 'merged'
> change was the patch change from a specific commit fom somewhere else
> (which may be missing from the local repo).
>
> Perhaps it needs a style of merging-rebase where a second (last) parent
> is added but it isn't the straight <sha1>, but says 'patch-<sha1>', such
> that readers with the capability could check if that <sha1> history is
> present locally, and if so if it's correct, so that you can now 'track'
> your fixes between releases, and (hopefully) older Gits don't barf on
> that extra 'fake' parent. Somehow I suspect that older Git's would
> barf.. (not enough time to create and test such a fake commit).

Sometime ago there was long discussion about adding 'weak' references to commit object header.

Beside the problem of backward compatibility, there was also the problem of semantics of said reference - what does it mean? It should work as well for cherry-picks, for interactive rebase (maybe?), and for reverts (which are also a problem).

Also, this could be avoided by using feature branches and merging instead of committing to one branch and cherry-picking to other branches. Also, git-rerere is your friend... sometimes.

>>> Have you tried git-imerge?
>>
>> No, not yet.  I’m not as interested in using it, as I would like git
>> itself to just work.

Maybe this command would make it into git proper, though probably not written in Python (there was once merge strategy written in Python, but currently git does not depend on Python).

-- 
Jakub Narębski