This doc explains the process and best practices for submitting a pull request to the Kubernetes project and its associated subrepositories. It should serve as a reference for all contributors, and be useful especially to new and infrequent submitters.
This guide is for contributors who already have a pull request to submit. If you’re looking for information on setting up your developer environment and creating code to contribute to Kubernetes, see the development guide.
First time contributors should head to the Contributor Guide to get started.
Make sure your pull request adheres to our best practices. These include following project conventions, making small pull requests, and commenting thoroughly. Please read the more detailed section on Best Practices for Faster Reviews at the end of this doc.
You can run these local verifications before you submit your pull request to predict the pass or fail of continuous integration.
make verify(can take 30-40 minutes)
Merging a pull request requires the following steps to be completed before the pull request will be merged automatically.
The Kubernetes merge workflow uses labels, applied by commands via comments. These will trigger actions on your pull request. Different Kubernetes repositories may require different labels on the path to approval. A generic explanation of how labels are used in pull requests can be found here. The pull request bot will also automatically apply and/or suggest labels.
Example: To apply a SIG label, you would type in a comment:
NOTE: For pull requests that are in progress but not ready for review,
prefix the pull request title with
[WIP] and track any remaining TODOs
in a checklist in the pull request description.
Here’s the process the pull request goes through on its way from submission to merging:
@k8s-ci-robot assigns reviewers
If you’re not a member of the Kubernetes organization, a Reviewer/Kubernetes Member checks that the pull request is safe to test. If so, they comment
/ok-to-test. Pull requests by Kubernetes organization members do not need this step. Now the pull request is considered to be trusted, and the pre-submit tests will run:
/retestto rerun failed tests
Reviewer suggests edits
Push edits to your pull request branch
Repeat the prior two steps as needed until reviewer(s) add
(Optional) Some reviewers prefer that you squash commits at this step
Follow the bot suggestions to assign an OWNER who will add the
/approve label to the pull request
Once the tests pass, all failures are commented as flakes, or the reviewer adds the labels
/approved, the pull request enters the final merge queue. The merge queue is needed to make sure no incompatible changes have been introduced by other pull requests since the tests were last run on your pull request.
The GitHub “munger” submit-queue plugin will manage the merge queue automatically.
/test all [submit-queue is verifying that this pull request is safe to merge]
cncf-cla: yeslabel added to pull request)
/retestif the pull request is trusted
That’s the last step. Your pull request is now merged.
If you want to solicit reviews before the implementation of your pull request is complete, you should hold your pull request to ensure that the merge queue does not pick it up and attempt to merge it. There are two methods to achieve this:
/hold cancelcomment commands
[WIP]prefix to your pull request title
The GitHub robots will add and remove the
do-not-merge/hold label as you use the comment commands and the
do-not-merge/work-in-progress label as you edit your title. While either label is present, your pull request will not be considered for merging.
The commands doc contains a reference for all comment commands.
The Kubernetes developer community uses a variety of automation to manage pull requests. This automation is described in detail in the automation doc.
The end-to-end tests will post the status results to the pull request. If an e2e test fails,
@k8s-ci-robot will comment on the pull request with the test history and the
comment-command to re-run that test. e.g.
The following tests failed, say /retest to rerun them all.
Pull requests older than 90 days will be closed. Exceptions can be made for pull requests that have active review comments, or that are awaiting other dependent pull requests. Closed pull requests are easy to recreate, and little work is lost by closing a pull request that subsequently needs to be reopened. We want to limit the total number of pull requests in flight to: * Maintain a clean project * Remove old pull requests that would be difficult to rebase as the underlying code has changed over time * Encourage code velocity
A few factors affect how long your pull request might wait for review.
If it’s the last few weeks of a milestone, we need to reduce churn and stabilize.
Or, it could be related to best practices. One common issue is that the pull request is too big to review. Let’s say you’ve touched 39 files and have 8657 insertions. When your would-be reviewers pull up the diffs, they run away - this pull request is going to take 4 hours to review and they don’t have 4 hours right now. They’ll get to it later, just as soon as they have more free time (ha!).
There is a detailed rundown of best practices, including how to avoid too-lengthy pull requests, in the next section.
But, if you’ve already followed the best practices and you still aren’t getting any pull request love, here are some things you can do to move the process along:
Make sure that your pull request has an assigned reviewer (assignee in GitHub). If not, reply to the pull request comment stream asking for a reviewer to be assigned. This is done via a bot command (the bot may have suggestions for this) and looks like this:
Ping the assignee (@username) on the pull request comment stream, and ask for an estimate of when they can get to the review.
Ping the assignee on Slack. Remember that a person’s GitHub username might not be the same as their Slack username.
Ping the assignee by email (many of us have publicly available email addresses).
If you’re a member of the organization ping the team (via @team-name) that works in the area you’re submitting code.
If you have fixed all the issues from a review, and you haven’t heard back, you should ping the assignee on the comment stream with a “please take another look” (
PTAL) or similar comment indicating that you are ready for another review.
Read on to learn more about how to get faster reviews by following best practices.
Most of this section is not specific to Kubernetes, but it’s good to keep these best practices in mind when you’re making a pull request.
You’ve just had a brilliant idea on how to make Kubernetes better. Let’s call that idea Feature-X. Feature-X is not even that complicated. You have a pretty good idea of how to implement it. You jump in and implement it, fixing a bunch of stuff along the way. You send your pull request - this is awesome! And it sits. And sits. A week goes by and nobody reviews it. Finally, someone offers a few comments, which you fix up and wait for more review. And you wait. Another week or two go by. This is horrible.
Let’s talk about best practices so your pull request gets reviewed quickly.
Are you sure Feature-X is something the Kubernetes team wants or will accept? Is it implemented to fit with other changes in flight? Are you willing to bet a few days or weeks of work on it?
It’s better to get confirmation beforehand.
When you want to make a large or otherwise significant change, you should follow the Kubernetes Enhancement Proposal process.
Even for small changes, it is often a good idea to gather feedback on an issue you filed, or even simply ask in the appropriate SIG’s Slack channel to invite discussion and feedback from code owners. Here’s a list of SIGs.
Small commits and small pull requests get reviewed faster and are more likely to be correct than big ones.
Attention is a scarce resource. If your pull request takes 60 minutes to review, the reviewer’s eye for detail is not as keen in the last 30 minutes as it was in the first. It might not get reviewed at all if it requires a large continuous block of time from the reviewer.
Breaking up commits
Break up your pull request into multiple commits, at logical break points.
Making a series of discrete commits is a powerful way to express the evolution of an idea or the different ideas that make up a single feature. Strive to group logically distinct ideas into separate commits.
For example, if you found that Feature-X needed some prefactoring to fit in, make a commit that JUST does that prefactoring. Then make a new commit for Feature-X.
Strike a balance with the number of commits. A pull request with 25 commits is still very cumbersome to review, so use judgment.
Breaking up Pull Requests
Or, going back to our prefactoring example, you could also fork a new branch, do the prefactoring there and send a pull request for that. If you can extract whole ideas from your pull request and send those as pull requests of their own, you can avoid the painful problem of continually rebasing.
Kubernetes is a fast-moving codebase - lock in your changes ASAP with your small pull request, and make merges be someone else’s problem.
Multiple small pull requests are often better than multiple commits. Don’t worry about flooding us with pull requests. We’d rather have 100 small, obvious pull requests than 10 unreviewable monoliths.
We want every pull request to be useful on its own, so use your best judgment on what should be a pull request vs. a commit.
As a rule of thumb, if your pull request is directly related to Feature-X and nothing else, it should probably be part of the Feature-X pull request. If you can explain why you are doing seemingly no-op work (“it makes the Feature-X change easier, I promise”) we’ll probably be OK with it. If you can imagine someone finding value independently of Feature-X, try it as a pull request. (Do not link pull requests by
# in a commit description, because GitHub creates lots of spam. Instead, reference other pull requests via the pull request your commit is in.)
Put changes that are unrelated to your feature into a different pull request.
Often, as you are implementing Feature-X, you will find bad comments, poorly named functions, bad structure, weak type-safety, etc.
You absolutely should fix those things (or at least file issues, please) - but not in the same pull request as your feature. Otherwise, your diff will have way too many changes, and your reviewer won’t see the forest for the trees.
Look for opportunities to pull out generic features.
For example, if you find yourself touching a lot of modules, think about the dependencies you are introducing between packages. Can some of what you’re doing be made more generic and moved up and out of the Feature-X package? Do you need to use a function or type from an otherwise unrelated package? If so, promote! We have places for hosting more generic code.
Likewise, if Feature-X is similar in form to Feature-W which was checked in last month, and you’re duplicating some tricky stuff from Feature-W, consider prefactoring the core logic out and using it in both Feature-W and Feature-X. (Do that in its own commit or pull request, please.)
In your code, if someone might not understand why you did something (or you won’t remember why later), comment it. Many code-review comments are about this exact issue.
If you think there’s something pretty obvious that we could follow up on, add a TODO.
Read up on GoDoc - follow those general rules for comments.
Nothing is more frustrating than starting a review, only to find that the tests are inadequate or absent. Very few pull requests can touch code and NOT touch tests.
If you don’t know how to test Feature-X, please ask! We’ll be happy to help you design things for easy testing or to suggest appropriate test cases.
Your reviewer has finally sent you feedback on Feature-X.
Make the fixups, and don’t squash yet. Put them in a new commit, and re-push. That way your reviewer can look at the new commit on its own, which is much faster than starting over.
We might still ask you to clean up your commits at the very end for the sake of a more readable history, but don’t do this until asked: typically at the point where the pull request would otherwise be tagged
Each commit should have a good title line (<70 characters) and include an additional description paragraph describing in more detail the change intended.
General squashing guidelines:
Do squash when there are several commits to fix bugs in the original commit(s), address reviewer feedback, etc. Really we only want to see the end state and commit message for the whole pull request.
Don’t squash when there are independent changes layered to achieve a single goal. For instance, writing a code munger could be one commit, applying it could be another, and adding a precommit check could be a third. One could argue they should be separate pull requests, but there’s really no way to test/review the munger without seeing it applied, and there needs to be a precommit check to ensure the munged output doesn’t immediately get out of date.
A commit, as much as possible, should be a single logical change.
Sometimes we need to remind each other of core tenets of software design - Keep It Simple, You Aren’t Gonna Need It, Minimum Viable Product, and so on. Adding a feature “because we might need it later” is antithetical to software that ships. Add the things you need NOW and (ideally) leave room for things you might need later - but don’t implement them now.
Sometimes reviewers make mistakes. It’s OK to push back on changes your reviewer requested. If you have a good reason for doing something a certain way, you are absolutely allowed to debate the merits of a requested change. Both the reviewer and reviewee should strive to discuss these issues in a polite and respectful manner.
You might be overruled, but you might also prevail. We’re pretty reasonable people. Mostly.
Another phenomenon of open-source projects (where anyone can comment on any issue) is the dog-pile - your pull request gets so many comments from so many people it becomes hard to follow. In this situation, you can ask the primary reviewer (assignee) whether they want you to fork a new pull request to clear out all the comments. You don’t HAVE to fix every issue raised by every person who feels like commenting, but you should answer reasonable comments with an explanation.
No document can take the place of common sense and good taste. Use your best judgment, while you put a bit of thought into how your work can be made easier to review. If you do these things your pull requests will get merged with less friction.