Easy Non-Fast-Forward Pushes

As far as I understand, a push will always modify (or add) a ref in the remote repository. When pushing to branch B, then the ref pointing to the last commit in this branch will be moved, provided that this can be done in a fast-forward way. Otherwise the push will fail.

The following options exist:

(1) force the push (2) fetch, merge, then push (3) push to a different branch

I don't want (1), for obvious reasons.

Option (2) implies that a particular person has to do the merge, namely the
person for whom the push is failing, or in other words: the person
unfortunate enough not being the first to push. (This reminds me of
Subversion: whoever tries to commit after someone else has committed will
have the burden of an update and possible merge.)

Option (3) allows others to recognize the situation, and anyone with repository access can do a merge. This is a good thing. However, I am confused as to what branch name should be used for this. In Mercurial, we would say that a new "head" is created, and anyone can recognize this by using "hg heads" and do a merge if he feels competent enough. (A "head" in Mercurial is a revision without children.) Can something similar be done in Git? I'd like to provide my co-workers with a command which they can always use to push their changes to a central repository in order that I can see what has happened and do merges accordingly. In Mercurial, such a command would simply be "hg commit && hg push -f". Then I can use "hg heads" to see if any new heads have grown, and if so, do merges. Upon their next pull, these merges would manifest themselves in the repositories of my co-workers, and everything will be fine for them. So, im looking for the equivalent of that workflow in Git. Thanks a lot!

On Fri, Feb 27, 2015 at 8:20 AM, Lasse Kliemann <lasse@lassekliemann.de> wrote:

For a high level workflow with pure git, everything must be a branch or a tag IMHO, (I cannot really think of other ways).

So maybe you create a bash alias for alias gitup='git push origin HEAD:${USER}/$(date -Iseconds)' which would push your current tip of the repository to the remote with quite a unique name.

Then you could also do a "git commit -a && gitup" to push your changes
to the server
As the integrator you could then integrate branches with
    "git fetch origin && git merge origin/sbeller/2015-02-27T09:34:47-0800"

So it is doable. Though I am not convinced of the workflow. Maybe you want to look at Gerrit or GitLab (both are open source web userinterfaces, where you can push changes to and approve them)

Stefan Beller <sbeller@google.com> writes:

This solution is not too bad, I thought, but there should be a way to remove those branches once I have integrated them. Otherwise the repository might get cluttered, and I found that those branches are particularly annoying when viewing a graphical representation of the commit graph.

Deleting is possible, using 'git push origin :BRANCH'. But then, anyone with pushing capabilities could do this! I don't like the possibility of one team member messing up another's commit, be it on purpose or accidentally.

In Gitlab, branches can be protected. For this to work, however, I would need fixed names. So I would assign one branch to each team member, make those branches (let's call them "personal branches" for now) protected, and then configure their systems (by shell aliases, scripts, ...) to the following:

1. Try pushing to origin/master. If it works, fine. If not, goto 2.

2. Push to the appropriate personal branch.

This system does not prevent someone pushing to another's personal branch, thereby possibly making a fast-forward push impossible. So step number 2 may fail due to this, but this might be an acceptable risk.

Thanks a lot for your thoughts.

Lasse Kliemann <lasse@lassekliemann.de> writes:

I wonder what happens to this user _after_ that change gets integrated on the project side. Presumably somebody picks up the change from the "personal branch", does necessary merge and updates the master, so the next time "git pull" is done, it will fast-forward?

I have a feeling that running trivial merges on the server-side when a push is made, and immedately pulling that result back might help such userbase who does not care too much about the history better, instead of using the bare-metal 'git pull' and 'git push'. You'd be scripting on the client side to do the above two steps for your end users anyway, so it would not be too much of a stretch to make that script a bit smarter still?

Lasse Kliemann <lasse@lassekliemann.de> writes:

I do not think I agree with the premise of the above in the first place.

The person whose push is failing is not a "victim". Everybody who pushes is working together with others to advance what is hosted in the repository he or she is pushing into.

If your work you failed to push is independent from others, then "fetch-merge-push" will be conflict-less and that is no basis to complain that you had to do extra work to get your change to the commons.

If your work does conflict, what does it mean? It means that other people did some work in the same area as you just have worked on. Who are the best equipped to resolve such a conflict?

True, you are not the only one who can understand the conflict. Those whose work in the same area that caused the conflict may know the nature of their own changes just as well as you know your changes, so they are equally qualified as you are. But you are among the people a lot better qualified to understand and resolve this conflict correctly than a random joe who was not involved in changes that happened during the time your work was done offline while the shared commons were advancing.

Unlike your view to see conflict resolution is a "burden", implying "after I push my changes out, it is other people's problem to correctly integrate their changes, and it will not be my problem if they break the change I did before they did with the merges", the workflow Git suggests when it rejects a non-fast-forward push encourages more effective collaboration by making sure that the opportunity to integrate _your_ own changes correctly to the shared common is given to _you_.

Having said all that.

There may be cases where you are far less qualified to resolve the conflict than other people who were involed in it. After

    $ git push
    ... results in no-ff
    $ git pull 
    ... results in conflict

the changes that overlap with your work can be found in this:

    $ git log ..MERGE_HEAD

And those who worked on them may know the area you and they both worked on a lot better than you. And I fully understand and agree with you that you and they need a way to communicate and makes it easy for you to ask them for help.

I however do not think an anonymous "split head" (i.e. 'master' that is split 47 ways because different people punted on their conflict resolution and left their work based on older state not usable by the collective commons) is the best way to do so. Your option (3) encourages a workflow where it is OK to punt after saying "I gave random others to resolve if/when they feel like it", while making it more difficult for you to communicate those others that are more qualified to help you---you have to say which one of 47 split heads you want them to take a look to help you.

Your message to ask somebody for help would say what your changes were for, what changes from them conflicted with that, and because you do not understand some of their changes you do not know how best to adjust your changes on top of theirs (yes, adjusting your changes to the updated shared commons is what a merge resolution is about). As a part of that communication, pushing your work, that is based on an older shared common from their point of view, to a branch whose name is taken after the topic you were working on, i.e. what your changes were for, would help (or a non-branch ref, such as refs/helpme/lasse-<topic>) a lot better than "Hey, here is a random unexplained half-done work based on an older version of 'master'. Feel free to grab it and integrate. I don't bother understanding and resolving the conflicts with more recent state of the project", which is how Option (3) seems to be aiming at.