Kai Wang (kaiwang@) and I (Jói Sigurðsson, joi@) experimented with something we called “shift-based development” in the first half of Q1 2013 as a way to work closely together on a componentization that was hard to do in parallel.
I work from Iceland, which is 7 or 8 hours ahead of Kai’s location (Mountain View, California), depending on the time of year (daylight savings time is not observed in Iceland). Kai and I were working on componentizing the Prefs subsystem of Chromium, and it was obvious early on that if we tried to develop in parallel, we would step on each others’ toes very regularly and be forced to do often difficult merges (due to e.g. moving files, renaming things, and so on). The idea came up, since my normal work day ends right around the time Kai’s starts, why not do the development in serial instead. Although I often work an extra hour or two later in the evening, that’s normally for responding to code review requests and email and not so much for coding, so this way we’d be completely free of getting in each others’ way.
The way we implemented this was we set up a bare git repository, and at the end of the day, we would push whatever we were working on to this repository, and let the other know the status of things. This could be a single working branch, or (more often) a pipeline of branches, plus a branch representing the SVN revision we were based off of. To make this work, we were both using the unmanaged git workflow.
The idea was, the next “shift” would pull down the pipeline of branches and continue where the last left off, doing development on the last change in the pipeline if it was incomplete, landing whatever changes could be landed, or starting a new change in the pipeline if all of them were ready for review or ready to land.
One limitation we ran into was that only the owner of an issue in Rietveld could upload a new patch set. To work around this limitation, I added a feature to Rietveld where you can add a COLLABORATORfirstname.lastname@example.org line to your change description, which will allow that person to also upload patches and edit the change description (see Rietveld patch).
In my opinion this was moderately successful. We were probably less productive than we would have been if each of us had been working on completely unrelated things, but certainly more productive than if we had tried to work together on componentizing Prefs in parallel.
With more practice, I think this way of working together could be quite successful. It was also challenging and fun and could be a worthwhile thing to try for folks separated by close to a full working day or more. See details below if you'd like to try.
The following instructions assume Linux is being used, but should be easily adaptable to other OSes. If you find mistakes in the instructions, please feel free to correct and clarify this Wiki page.
On one of our Linux boxes, we set up a bare git repository using these instructions. I'm sure you could also use an existing git service such as github.
Let's assume the IP address of the Linux box hosting the bare repository is
220.127.116.11, the Linux user that provides access to the bare repository is named
gitshift, and the repository is located at
Each developer participating in the shift-based development provides their RSA public key (e.g.
~/.ssh/id_rsa.pub) and we add it to
/home/gitshift/.ssh/authorized_keys; that way as long as you are using
ssh-agent and have run
ssh-add, you won't need to type in a password every time you issue a git command that affects the repository.
Issue these commands to add a remote for this repository to your local git repo, and to mirror its initial state. You can call it whatever you like, shiftrepo is just an example.
$ git remote add shiftrepo email@example.com:/home/gitshift/git/gitshift.git $ git fetch shiftrepo
You should now be able to do a
git push of some dummy branch; try it out, e.g.
$ git checkout -b shifttest master $ echo boo > boo.txt $ git add boo.txt && git commit -m . $ git push shiftrepo
The shared repository is just a place where you share your branches; it is not a place where you do actual work. The actual work should be done in your separate local repositories, and you still use e.g. git pull (which in our git svn repositories behind the scenes does a
git svn fetch etc.
Shift-based collaboration won't work well (at least not with a pipeline of branches) unless you are using an “unmanaged” git checkout.
For the branches we collaborated on, we set up some working rules. This may be a good starting set and it worked well enough for us, but others could adapt these rules:
a) We had a naming convention for pipelines of branches. E.g. you might have branches named shiftrepo/p0-movemore and shiftrepo/p1-sync, where p1-sync‘s upstream branch is p0-movemore, and p0-movemore’s upstream branch is an SVN revision, generally a version that was LKGR at some point. You can find the git commit hash of this revision by running
git log -1 --grep=src@ | head -1 | cut -d " " -f2
and the SVN revision number by running
git log -1 --grep=src@ | grep git-svn-id | cut -d@ -f2 | cut -d " " -f1
We followed a naming convention of one or two alphabetic characters followed by the sequence number of the branch, followed by a dash and a descriptive name for what's going on in that particular branch. The one or two alphabetic characters indicated the rough over-arching topic (e.g. p for Prefs), and the stuff after the dash can be more descriptive.
b) When pushing a pipeline of branches to shiftrepo (where branch A depends on branch B and so forth) we made sure to first git pull in each dependent branch in sequence, so that each branch in shiftrepo is building straight on top of the previous branch.
c) At the end of our shift, we did
git push shiftrepo branchname for each branch.
d) At the start of our shift, we did
git fetch shiftrepo and then for each branch we were collaborating on we did
git checkout branchname && git merge shiftrepo/branchname. Note that the first command checks out the local branch, and the second merges the shiftrepo/ branch into it. This does not make the shiftrepo/ branch a parent of the local branch.
e) Also at the start of each shift, we updated the local upstream branches for each branch to match the upstream relationships that the person ending their shift had on his end. One case is if the oldest branch in the pipeline has been merged to a new LKGR, then we did this:
git branch --set-upstream oldestBranchName `git log -1 --grep=src@ oldestBranchName | head -1 | cut -d " " -f2`
and the other case is if new branches were created during the last shift, e.g. p4-foo was added, then for each we need to do like this:
git branch --set-upstream p4-foo p3-bar
f) For managing old branches, we removed the oldest branch in a pipeline when several conditions were met:
i) The old branch has been checked in.
ii) The SVN revision of the check-in of the old branch is equal to or older than LKGR, i.e. that change in SVN is included when you sync to LKGR.
iii) That LKGR has been merged into the old branch, and we've done
git pullin the next branch after it.
g) We only used the CQ to commit stuff. The fear was (and this hasn't really been validated as true or false) that there might be some gotchas if we used
git cl dcommit instead.
h) At the end of our shift, we communicated by email/IM/Hangout to let the other know the status of the work, next steps remaining for any currently-open branches, and to discuss what might make sense for the next branches to work on.
To push a local branch to shiftrepo:
git push shiftrepo localbranchname
To push all “matching” branches (i.e. push the latest copy of any local branch that has previously been pushed to shiftrepo):
git push shiftrepo
To delete a branch from shiftrepo, it's weird:
git push shiftrepo :branchname