We follow the Git project's own commit discipline practice of "Each commit is a minimal coherent idea". This discipline takes a bit of work, but it makes it much easier for code reviewers to spot bugs, and makes the commit history a much more useful resource for developers trying to understand why the code works the way it does, which also helps a lot in preventing bugs.
Commits must be coherent:
- It should pass tests (so test updates needed by a change should be in the same commit as the original change, not a separate "fix the tests that were broken by the last commit" commit).
- It should be safe to deploy individually, or explain in detail in the commit message as to why it isn't (maybe with a [manual] tag). So implementing a new API endpoint in one commit and then adding the security checks in a future commit should be avoided -- the security checks should be there from the beginning.
- Error handling should generally be included along with the code that might trigger the error.
- TODO comments should be in the commit that introduces the issue or the functionality with further work required.
Commits should generally be minimal:
- Significant refactorings should be done in a separate commit from functional changes.
- Moving code from one file to another should be done in a separate commits from functional changes or even refactoring within a file.
- 2 different refactorings should be done in different commits.
- 2 different features should be done in different commits.
- If you find yourself writing a commit message that reads like a list of somewhat dissimilar things that you did, you probably should have just done multiple commits.
When not to be overly minimal:
- For completely new features, you don't necessarily need to split out new commits for each little subfeature of the new feature. E.g., if you're writing a new tool from scratch, it's fine to have the initial tool have plenty of options/features without doing separate commits for each one. That said, reviewing a 2000-line giant blob of new code isn't fun, so please be thoughtful about submitting things in reviewable units.
- Don't bother to split backend commits from frontend commits, even though the backend can often be coherent on its own.
- Overly fine commits are easy to squash later, but not vice versa. So err toward small commits, and the code reviewer can advise on squashing.
- If a commit you write doesn't pass tests, you should usually fix that by amending the commit to fix the bug, not writing a new "fix tests" commit on top of it.
Zulip expects you to structure the commits in your pull requests to form
a clean history before we will merge them. It's best to write your
commits following these guidelines in the first place, but if you don't,
you can always fix your history using
git rebase -i.
Never mix multiple changes together in a single commit, but it's great to include several related changes, each in their own commit, in a single pull request. If you notice an issue that is only somewhat related to what you were working on, but you feel that it's too minor to create a dedicated pull request, feel free to append it as an additional commit in the pull request for your main project (that commit should have a clear explanation of the bug in its commit message). This way, the bug gets fixed, but this independent change is highlighted for reviewers. Or just create a dedicated pull request for it. Whatever you do, don't squash unrelated changes together in a single commit; the reviewer will ask you to split the changes out into their own commits.
It can take some practice to get used to writing your commits with a clean history so that you don't spend much time doing interactive rebases. For example, often you'll start adding a feature, and discover you need to do a refactoring partway through writing the feature. When that happens, we recommend you stash your partial feature, do the refactoring, commit it, and then unstash and finish implementing your feature.
The first line of the commit message is the summary. The summary:
- is written in the imperative (e.g., "Fix ...", "Add ...")
- is kept short, while concisely explaining what the commit does
- is clear about what part of the code is affected -- often by prefixing with the name of the subsystem and a colon, like "zjsunit: ..." or "docs: ..."
- is a complete sentence, ending with a period.
zjsunit: Fix running stream_data and node tests individually.
gather_subscriptions: Fix exception handling bad input.
Add GitLab integration.
Compare "gather_subscriptions: Fix exception handling bad input." with:
- "gather_subscriptions was broken", which doesn't explain how it was broken (and isn't in the imperative)
- "Fix exception when given bad input", in which it's impossible to tell from the summary what part of the code is affected
- "gather_subscriptions: Fixing exception when given bad input.", not in the imperative
- "gather_subscriptions: Fixed exception when given bad input.", not in the imperative
The summary is followed by a blank line, and then the body of the commit message.
- The body is written in prose, with full paragraphs.
- The body explains:
- why and how the change was made
- any manual testing you did in addition to running the automated tests
- any aspects of the commit that you think are questionable and you'd like special attention applied to.
- If the commit makes performance improvements, you should generally include some rough benchmarks showing that it actually improves the performance.
- When you fix a GitHub issue, mark that you've fixed the issue in your commit message so that the issue is automatically closed when your code is merged. Zulip's preferred style for this is to have the final paragraph of the commit message read e.g. "Fixes: #123."
- Any paragraph content in the commit message should be line-wrapped
to less than 76 characters per line, so that your commit message
will be reasonably readable in
git login a normal terminal.