- I've been using the "apply mail" flow to avoid merge commits and to squash commits when appropriate (based on Nathaniel Talbott's post http://blog.spreedly.com/author/ntalbott/#.VGVhz5PF_Zs). It will be nice to have something automatic.
- At least contributors with the commit bit can just create a feature branch and a pull request from there. Some open source projects give out the commit bit almost immediately, but with GitHub there's still that initial fork and pull request to prove oneself.
Interested to see how Gerrit addresses this. GitHub doesn't have any way to disable pull requests. :-(
- I've completely disabled email notifications for this reason. There still can be a huge number of notifications in app or via third-party mobile apps. Rietveld is a lot more sane, I only saw notifications for things I was actively working on.
- I usually end up reviewing just the new commits individually. But that requires figuring out which ones I had already reviewed. Not so bad with only a few branches, but I can see it getting out of hand.
I really hope people from GitHub are reading this thread as they are working to improve their tools. :-)
If not then it just seems like switching to git using the present google code host would suffice from a technical viewpoint.
It seems like it must be possible to hook up a bot to github that watches for pull requests and submits them to gerrit... the Juju team has a bot that does that for our Reviewboard integration. It won't work for more complicated workflows that github doesn't support, like dependent branches or whatever, but it lowers the barrier of entry for initial contributors that just want to submit a simple fix. And the more advanced users can just submit directly to our Reviewboard instance if they want the more advanced features.
Also, the public fork on Github is actually a feature - it means others can easily see what you're working on. You need off-site storage for your personal work anyway, right? You're not going to keep it local on your laptop, so you'll need a branch somewhere regardless.
You can create Pull Requests within a repository. You don't need to fork the repository if you have push access.
1 and 3 have their usecases, 3 is esp. useful if the branch contains only 1-2 commits, and 1 is useful if the branch contains > 10. You can add test information into the merge commit.
I hope github will take more gerrit features into their tracker or allow tracker extensions, so that we can live with github alone. gerrit still is a UI and performance nightmare.
For Go, we will require that contributors (everyone working on Go) prepare their commit at head (or at least make the commit able to be automatically rebased). It makes no sense to have a merge commit for every single commit. That would add no extra information and double the size of the commit log.
What's the purpose of keeping a commit history at all? Really, what are your use cases? For me, the history is for looking around and understanding a particular change, or for bisecting to help understand a bug. Both of these use cases are better supported by seeing the tree as it actually was when someone was working on it, than by seeing an artificial, rebased history. What's your use case where the rebased version makes sense?
- Merging a pull request (almost) always creates a merge commit, polluting the change history. Gerrit will automatically rebase the changes atop the master branch head, leaving a nice linear history.
- Pull requests require the contributor to create a public fork of the repository they're committing to. I can see how this works for some people, but I find it gross for each contributor to the Go project to have their own public fork. What a mess.
- Comments on pull requests are sent as soon as they are created. Gerrit allows you to make many comments on a change, as drafts, and then send them all in one go, sending just a single email. This is much easier to manage for a large project.
- Gerrit has the notion of multiple 'patch sets' for a particular change, and you can see diffs between patch sets, so it's much easier to progressively review large changes.
And there are many other small issues with the GitHub pull request process that make it untenable for the Go project.