So we already use a internal refspec for tracking the latest PR HEAD. You can actually manually fetch this under `refs/pull/123/head`. However, this is only the latest HEAD, not any previous histories.
My idea was to expand this to include tracking all historical HEAD refs. We came up with a clever/hacky idea to string together a chain of dummy commits pointing at all the historical HEADs. This fake commit chain would only be used internally, but it would ensure git's gc reachability checks don't sweep up any old PR HEADs. Its an alternative to creating a new ref for each previous HEAD or maintaining text reflogs for an entire repo network. Both have some performance issues for larger repos.
Its something we're still working on, but it still helps to vocalize your support for wanting improved force push support. :D
YES! I wish for that feature everyday. Gerrit does it exactly right.
Related to this, I really, really hope that you will consider displaying commits in a PR in the correct order, suppose that I have three commits on top of master:
a->b->c->master
If I git rebase -i master and reorder my commits so that they look like this: c->a->b->master
Then the commit list in the PR will still be displayed in chronological order
instead of DAG order, this is very confusing! The only workaround I've found is
to amend every commit in my PR so that the timestamp order match the DAG order,
this isn't very fun to do.By the way, the post explicitly says:
> Some teams choose to use a commit-by-commit workflow where each commit is treated as the unit of change and is isolated for review.
I'm glad that you're acknowledging that. But then why do comments on commits still get hidden when they correspond to an outdated diff? I would love to be able to comment on individual commits, but if half my comments are hidden because they correspond to lines that don't match the current diff, then I'm not going to do it for fear of my comments being missed. You should not hide important information that I'm trying to communicate! Here's a better way to do it: Add a "Done" button like Rietveld/Gerrit for every comment (this is a very useful feature on its own) and hide comments _if and only if_ the "Done" button has been clicked.
Anyway, I'm glad that people are working on this stuff and that they're listening to the community, thank you!
"ok, I fixed that, but what do you think about xyz?"
Now I have to go to the comments and expand 35 of them until I find the right one so I can respond to their question.
It's also sometimes the case that changing a line is not the same as addressing a comment, or maybe it only addresses part of a comment, or maybe they just changed whitespace because now all that code you commented on is inside an if...
Example in action: https://github.com/grpc/grpc-go/pull/570
unfortunately, bots like these are very annoying to write right now. at least one reason is that there is no github event for editing comments. if you try to use webhook events to mirror comments made in github to another code review tool, you will quickly drift out of sync if those comments are edited in github. instead, you have to implement polling, which comes with it's own set of challenges.
also, in general, the event formats are pretty inconsistent. as a random example, in the push event the repo owner only includes name and email, while in the commit comment event the repo owner includes all sorts of information about stars, followers, etc... there are inconsistencies like that all over the place, which means you might need to implement additional API requests for some events and not others.
on a somewhat related note, i've seen a lot of speculation in the open-source community that the reason open-source is getting so little attention from Github is that they're focusing on enterprise customers. i can say that as a somewhat-large enterprise customer, i have yet to get a single issue addressed. examples like above (add a comment-edited and PR-edited event) have been in the ask 6 months or more.
The dummy commit approach you outlined sounds like a good idea. It might be nice to expose that externally similar to how `refs/pull/123/head` and `refs/pull/123/merge` work, so that way we can handle force-pushes to PRs better in our own tooling. I'd suggest something like `refs/pull/123/history`.
I don't always want to see my diff in alphabetical order.
I would love to have more clarity on release schedules for enterprise in general
Using this flow we never lose any comments during the PR process and can revisit the individual commits, diffs and files changed while the PR is open.
I encourage members on our team to rebase frequently as they're iterating prior to opening their PRs for their individual feature branches and I think this helps us avoid problems for our typical PRs.
Also as a rule, don't sign off on a PR until all conflicts are resolved, if there were any and review the final commit(s) carefully.
We also have the option because of some of the CI tools we've written to give our QA people access to specific SHAs that can be tested, so sometimes (rarely) QA sign-off is asked for as part of the PR.
http://fle.github.io/git-tip-keep-your-branch-clean-with-fix...
Specifically, I would love to see GitHub extend its searching capabilities, and its code review, to be able to encompass more of a Gerrit style approach.
For example, let's say you want to have two core reviewers give a positive review on a PR before it lands. Right now your best hope with GitHub is tags, and that is frail and complex. In Gerrit, it's just some configuration.
Or, being able to show yourself only PRs that you haven't reviewed yet, or maybe ones that haven't had any reviews yet. In gerrit, it's some search terms. In github, it's another big complex tag operation.
When you make a change, all your new commits show up in a new version, and you can switch between the different versions of the MR to see the old state of the code before each new version. Where a version is created after a push of new commits to the branch.
You still keep the entire commit history (which I much prefer), but the view of that history and discussion around it is much clearer in the MR.
Github's way of doing it makes it difficult to see a concise view of the history of the MR, which can contain very valuable information outside of the commit history.
I'll shortly be deploying a new feature where Reviewable will try to match up rebased commits with their predecessors and automatically pair them up when diffing. This is just a heuristic based on commit messages but it works pretty well for non-interactive rebasing. Otherwise, you can still manually diff any two revisions you'd like.
Disclosure: I built the tool.
Gerrit instead retains each rewrite of the "same" commit. It does this by requiring you to insert a "Change-Id: ..." footer into each commit message (it provides a repo hook to do this) and examines each incoming commit message to know whether that commit is a revision to a commit it's already seen.
In our project we exclusively use comments on the Files tab precisely for that reason.
The difference between commenting on the commit page and on the "files tab" is pretty subtle though...
1) piling new commits onto the existing branch/PR, or
2) force-pushing and completely losing the old commits on the server
You could instead push into a "magic" ref-spec and the server would retain the original commits, but also the rewritten commits, such that a PR can have multiple revisions. This is somewhat hard to explain unless you're familiar with Gerrit and its "patch set" workflow.
Why? Because my preference is to rewrite history in order to address code review comments, such that what is finally merged leaves a clean history. But it is also valuable during the code review process to be able to look across the revisions to make sure everything has been properly addressed.
The only way to do both, today, is to create a new branch and PR for each "revision". i.e. new-feature, new-feature-rev1, new-feature-rev2. Then close the original PR and reference it from the new PR. A bit tedious.