Hacker News Reader

Comment by sirn | Hacker News Reader

One thing to note if someone consider switching from GitHub to Phabricator: Phabricator use a pre-commit reviews based on patches rather than being branch-based like in GitHub/GitLab/BitBucket. Make sure to read the Phabricator Review Workflow[1] before making the switch, because everyone in your team have to change their workflow if they already used to Pull Request/Merge Request workflow. :-)

That said, I always found Phabricator's issue tracker (Maniphest) to be superior to GitHub Issues, with ability to create task dependencies (parent/child) and ability to limit tasks permissions only to few selected groups (with spaces). It's possible to setup Phabricator to mirror other repositories if you want to use only its issue tracker.

[1]: https://secure.phabricator.com/phame/post/view/766/write_rev...

View on HN

The patch-based workflow is very similar to how the Linux kernel review workflow works. Arguably, this is the way Git is supposed to work - submit patches, discuss them, then merge them once they're ready. Every idea is one commit, no checkpoints or "fix typo". Pull requests are used to merge other maintainer's trees that have many commits in them. It also ensures that there's a permanent link between a commit and its review, since there's a 1:1 mapping.

Gerrit works similarly, but Phabricator's UI is much more friendly.

For small projects or teams, the overhead of introducing a new tool may not be worth it, for for large projects with many contributors, the workflow is - in my opinion - strongly superior.

If you set Github to "squash and rebase" and enforce a "one PR per idea" rule, you will get a pretty similar workflow. The tooling isn't quite as polished, but it's sufficient for many projects.

Quote:

> As far as code reviews go, this is pretty spot on. I was part of a a startup that was using GitHub pull requests for code reviews. As the team grew, it became more and more intractable, although not simply because of notifications. Side-by-side diffs and checkpointed diffs (so that you can see what changed since the last round of review and whether/how your comments were addressed) are handled very poorly by GitHub. We ultimately switched to Phabricator, and while there was a little friction as folks got acquainted with the new tool, it made code reviews a much more pleasant process. Recently, I had to go through a full code review back on GH pull requests, and it felt like pulling teeth in comparison. They're fine for interacting with contributors to an open source project, but compared to working with a tool like Phabricator that's built for a code reviewer's workflow (and for teams of engineers working together on a project), they just don't hold a candle, in my opinion.

(from a previous discussion: https://news.ycombinator.com/item?id=7697132)

> Side-by-side diffs and checkpointed diffs (so that you can see what changed since the last round of review and whether/how your comments were addressed) are handled very poorly by GitHub.

I've heard that complaint a lot. How are these things handled poorly by GitHub? It's an honest question; I've only used GH and a bad, ad-hoc, poorly-thought-through Perforce workflow, so I may not know what I'm missing, but it has always seemed quite simple for me to scroll through a pull request and see feedback (either tagged to lines of diff or general), and commits that address pieces of feedback. Then you can click on each commit to view it, or open the full hairball diff to see how those changes end up in the final result. What's missing there?

It sucks when someone puts a bunch of unrelated changes in one commit, or when their commit message is bad ("code review"), but those seem like they'd be problems regardless of GitHub.

I've looked at patch-submit-discuss/mailing-list-style workflows, and they seem fine. The discussion is definitely more front and center. However, when people submit updated patches it seems like I'd want to be able to easily tell what changed between their first iteration and second iteration of the patch--i.e. did they actually address the feedback they got? This can be hard to see, especially if a patch is substantial or to confusing/repetitive code. I can download all the submissions and run diff on each pair in sequence, but that seems like needless hassle when, unless the feedback was "throw it all out and do it a totally different way" (which, in a branching workflow, is often a good reason to start on a fresh branch), subsequent iterations could be submitted as a "patch to a patch", with an easy way to view the cumulative diff. At this point, it would basically be a branch workflow. Are there solutions for this in the patch-submit-discuss model?

> I've heard that complaint a lot. How are these things handled poorly by GitHub?

Basically, when changes are requested, the patch based work flow is to submit a revised set of commits up for review. In Github, one can do that by either force-pushing to the branch that's currently attached to a pull request, or one can push up a new branch and open up a new pull request associated with that branch.

But there isn't a way to view the "interdiff" between the version of the branch before review and the version of the branch after review. Review tools like Phabricator, Gerrit, and email lists allow one do view that diff.

> But there isn't a way to view the "interdiff" between the version of the branch before review and the version of the branch after review. Review tools like Phabricator, Gerrit, and email lists allow one do view that diff.

Yes there is. You even get a link to view that exact diff whenever someone updates a PR. It shows changes since you last viewed the PR, up to its current state.

That's only the case if someone pushes up a commit like "Addressing comments" or "Fixed bug". If you rebase locally and force push the branch to maintain the clean set of commits after addressing review comments, Github does not provide a way to see the diff between the previous branch head commit and the current head commit.

> If you rebase locally and force push the branch to maintain the clean set of commits after addressing review comments

Why would you do that, in general? That seems needlessly hostile to reviewers, since then they have to go through your diff and make sure you caught every place their comment applied (e.g. "did they actually handle the else-case everywhere they're testing on $var?").

Is this common practice in situations* where it harms diff readability? If so, why?

* There are totally cases where that's necessary, but they don't seem like more than a small minority of comment/feedback/amendment cases. These are times like "you put all of this code in the totally wrong place" or "this is missing a fundamental assumption that would require you to totally rearchitect every change you wrote". In those cases, a rebase (plus clearly calling out "everything was rebased starting here please start reading from here on!") or a brand new branch is in order.

> Why would you do that, in general?

The main purpose is to maintain a clean set of commits. So rather than having a history like:

    Add method and unit tests
    Update version to 1.2.3
    Fixed issue in method
    Fixed syntax error
    Addressing review comments

you only have the first two commits.

> That seems needlessly hostile to reviewers

Why would doing something like this be considered hostile? The problem here is the tool used for reviewing code, not the submitter or reviewers. Other review tools have had the feature of viewing the diff of the branch across rebases for quite some time now.

The only other alternative is to rely on the submitter to make sure they squash the amendments they make into the correct commit. That in itself is more difficult after making multiple commits of that nature. What usually happens in that case is that those commits end up as part of the master branch. This makes it more difficult to revert certain changes because they're split across multiple commits and it also makes it more difficult to track down the source of a bug. It also makes it more difficult to see which commit was responsible for a given change when using git blame.

The problem I run into when reviewing PRs on Github is that many people rebase or otherwise mutate their history locally to address review comments, then force push to the branch being reviewed. Once they've done that, you've lost the original commits and can't see how the new ones differ.

https://github.com/isaacs/github/issues/997

Only if the branch has commits appended, not if it has modified commits.

Isn't modifying commits considered a poor practice in the general case?

I dislike when questions are worded like that, because it could be interpreted in two ways:

a) Given an arbitrary case, is there a chance that modifying commits is considered a poor practise in that case?

b) Is the belief that modifying commits is poor practise held by the population at large?

The answer to a) is yes, the answer to b) is no.

Some people prefer a clean history, where you have to explicitly request the steps to get to it if you want them. Others want the messy steps to be part of the history you show up-front.

In Gerrit you can diff against previous versions of the patch, right from the web-ui.

> Every idea is one commit

Ideally, a commit should be self contained and change only one thing. This means that a feature would consist of multiple commits.

Since later commits may depend one or more earlier commits (some of which may be in the same feature branch), how do tools that work on a commit based workflow handle those dependencies?

Most large teams avoid long-lived feature branches and instead try to get things reviewed and merged into master as soon as possible. It's better to have many small reviews than a few massive ones.

You usually try to avoid dependencies. If you can't avoid them, have a series of patches depend on each other.

Phabricator supports stacking patches. Here's a random patchset:

https://secure.phabricator.com/D19429

Scroll down to "Stack" - it will display the list of patches depending on each other:

https://i.imgur.com/WTl4B8x.png

Dependent changes are a nightmare in phabricator. I tried it once using arcanist and git, and it seemed to be confused about which patchset I wanted to change. After this experience, I talked to a few colleagues, and they basically said "yeah, just don't use dependent reviews"

> The patch-based workflow is very similar to how the Linux kernel review workflow works.

I don't think that is accurate. Or there kernel its very common to submit whole patch series (branches), and to rebase them on the latest kernel. Phabricator makes multi commit reviews possible but a total pain.

It works exactly the same with Phabricator - you submit a series of individual patches and have them depend on each other.

That sounds horrible (to me). I have always preferred the branch/pull request/merge workflow. I always have always felt strongly that the history be maintained so we can see how the code evolved when a feature was being worked on. We can often get insight into what things may have been tried but didn't work.

It took some time to get used after moving from Atlassian Stash (which works quite well for code reviews, imho) to Phabricator/Arcanist. Still consider the latter inferior to Atlassian tools.

Another thing to consider on this is the management that has to be done to arc. When using arc, it has to be setup in every new computer, and it becomes messy if you want to use Client Certs.

We had a big Wiki page for newcomers on how to setup arcanist, and it was not that funny.

Reply on news.ycombinator.com