Skip to content

Latest commit

 

History

History
355 lines (258 loc) · 14.9 KB

RELEASE.md

File metadata and controls

355 lines (258 loc) · 14.9 KB
Raw

Kubeflow Pipelines Release Process

Schedule

Kubeflow Pipelines has weekly patch releases and monthly minor releases. Patch releases only contain bug fixes, while minor releases have new features additionally.

Release Tags and Branches

Releases are tagged with tags like X.Y.Z, e.g. 1.0.2. A special format like 1.0.0-rc.2 is a pre-release. It is the second release candidate before releasing the final 1.0.0.

A release branch has the name structure release-X.Y where X.Y stands for the minor version. Releases like X.Y.Z will all be released from the branch release-X.Y.

For example, 1.0.2 release should be on release-1.0 branch.

Contributor Instructions

The following section targets contributors. No need to read further if you only want to use Kubeflow Pipelines.

Cherry picking pull requests to release branch

After the release-X.Y release branch is cut, pull requests (PRs) merged to master will be only get released in the next minor release X.(Y+1).0.

If you want your PR released earlier in a patch release X.Y.(Z+1):

  • The PR must already get merged to master branch.
  • The PR should be a bug fix.
  • The PR should be cherry picked to corresponding release branch release-X.Y.

Choose one of the following options for cherry picking your PR to release branch.

Option - (Recommended) cherrypick-approved label

Contributors should ask OWNERS who approved the PR to add a cherrypick-approved label if they want the PR cherry picked to release branch.

Release manager will periodically or before release, search all merged PRs with cherrypick-approved and cherry pick them into current release branch.

Option - git cherry-pick

  • Find the commit you want to cherry pick on master as $COMMIT_SHA.

  • Find the active release branch name $BRANCH, e.g. release-1.0

  • Cherry pick a commit:

    git checkout $BRANCH
git checkout -b <cherry-pick-pr-branch-name>
git cherry-pick $COMMIT_SHA

  • Resolve merge conflicts if any

  • git push origin HEAD

  • Create a PR and remember to update PR's destination branch to release-$MINOR_VERSION

  • Ask the same OWNERS that would normally need to approve this PR

Release Manager Instructions

The following sections target release managers. You don't need to read further if you only want to use or contribute to this repo.

Common Prerequisites

  • OS: Linux or MacOS
  • Permissions needed
    • Can create a branch in github.com/kubeflow/pipelines.
    • (Before #4840 is resolved) one would need the admin access to kubeflow/pipelines repo.
    • Can trigger cloudbuild jobs in google.com/ml-pipeline-test GCP project.
  • Tools that should be in your $PATH
    • docker
    • python3
  • Preparations
    1. Clone github.com/kubeflow/pipelines repo into $KFP_REPO.
    2. cd $KFP_REPO

Cutting a release branch (Optional)

  1. Choose a good commit on master branch with commit hash as $COMMIT_SHA.

  2. Choose the next release branch's $MINOR_VERSION in format x.y, e.g. 1.0, 1.1...

  3. Make a release branch of format release-$MINOR_VERSION, e.g. release-1.0, release-1.1. Branch from the commit and push to kubeflow pipelines upstream repo.

    git checkout $COMMIT_SHA
BRANCH=release-$MINOR_VERSION
git checkout -b $BRANCH
git push upstream HEAD

Before release

Do the following things before a release:

  1. (Do this step only when releasing from a NON-master release branch)

    Note: Instead of following this step to cherry pick all PRs, you can also manually cherry pick commits from master branch to release branch, if the number of PRs to cherry pick is minimal. Command for manual cherry pick:

    git cherry-pick <commit-id>

    If you want to use script to cherry pick all merged PRs with cherrypick-approved label:

    • Search all merged PRs with cherrypick-approved label, but no cherrypicked label using this link

    • Use the git cherry-pick option to pick these PR commits into the release branch one by one in a batch and add cherrypicked label to these PRs.

      NOTE: if there are merge conflicts for a PR, ask the PR author or area OWNER to create a cherry pick PR by themselves following other two options.

    • git push upstream $BRANCH directly to the release branch.

    There's an automated script that can help you do the above:

    # Prepare your env
cd ~/kubeflow/pipelines
git fetch upstream
git checkout release-1.0
git pull
git checkout -b <your-cherry-pick-branch-name>

# The following command shows usage info
./hack/cherry-pick.sh

# The following command cherry picks PRs #123 #456 #789 for you.
# It runs git cherry-pick for each merged commit, then adds `cherrypicked`
# label on the PR.
#
# If there's a merge conflict in the middle, it will stop there waiting for
# you to resolve. You need to add the `cherrypicked` label by yourself in
# this case. After the issue resolved, you can rerun the same command and
# PRs already cherrypicked (with the label `cherrypicked`) will be skipped.
./hack/cherry-pick.sh 123 456 789

# After cherry pickings are done, they are still in your local repo. Push
# them to your remote branch to create a PR.
git push origin HEAD

    You can get the list of PRs waiting to be cherrypicked by:

    1. Open cherrypick-approved PRs that haven't been cherrypicked sorted by updated order.

    2. Open browser console (usually by pressing F12).

    3. Paste the following command into the console.

          console.log(Array.from(document.querySelectorAll('[id^="issue_"][id*="_link"]')).map(el => /issue_(.*)_link/.exec(el.id)[1]).join(' '))

  2. Verify cloudbuild and postsubmit tests are passing: visit https://github.com/kubeflow/pipelines/commits/master for master branch.

How to very cloudbuild and postsubmit status

If not, contact the KFP team to determine if the failure(s) would block the release. You can also retry the failed job by opening the detail page of prow job, and click the refresh button next ot the job title.

Releasing from release branch

Note, when releasing from master, all the below mentions of "release branch" means master branch.

  1. Choose the release's complete $VERSION following semantic versioning, e.g.

    • 1.0.0-rc.1
    • 1.0.0-rc.2
    • 1.0.0
    • 1.0.1
    • 1.1.0
    • ... Set the version by using VERSION=<version-value>. Contact @chensun if you are not sure what next version should be.

  2. Release driver and launcher images

    First build and push driver and launcher images:

    pushd ./backend

export IMG_TAG_DRIVER=gcr.io/ml-pipeline/kfp-driver
make license_driver image_driver
docker push $IMG_TAG_DRIVER

export IMG_TAG_LAUNCHER=gcr.io/ml-pipeline/kfp-launcher
make license_launcher image_launcher
docker push $IMG_TAG_LAUNCHER

popd

    If there are changes to the licenses files, make a Pull Request and merge the changes. (e.g. #11177)

    Once the images are pushed, update the hard-coded hash in the code: DefaultLauncherImage and DefaultDriverImage. Make a Pull Request and merge the change (e.g.: #11178)

    Pull the changes:

    git checkout $BRANCH
git pull upstream $BRANCH

  3. Update all version refs in release branch by

    cd ./test/release && TAG=$VERSION BRANCH=$BRANCH make release

    It will prompt you whether to push it to release branch. Press y and hit Enter.

    Note, the script will clone kubeflow/pipelines repo into a temporary location on your computer, make those changes and attempt to push to upstream, so that it won't interfere with your current git repo.

    If you see error "docker.sock: connect: permission error", you need to allow managing docker as a non-root user.

  4. View related cloudbuild jobs' statuses by clicking the latest commit's status icon in the release branch. Refer to the screenshot below to find the page.

  5. Wait and make sure the build-each-commit cloudbuild job that builds all images in gcr.io/ml-pipeline-test succeeded. If it fails, please click "View more details on Google Cloud Build" and then "Retry".

    NOTE: you can find your latest release commit in https://github.com/kubeflow/pipelines/commits/master and select your release branch. How to very cloudbuild and postsubmit status

  6. Select the release-on-tag cloudbuild job that copies built images and artifacts to public image registry and gcs bucket. This job should have already failed because artifacts haven't been built. Now, please click "View more details on Google Cloud Build" and then "Retry", because after waiting for previous step, artifacts are now ready.

    NOTE: DO NOT click the "Re-run" button from GitHub Actions status page. It will create a build with "Branch: $BRANCH" instead of "TAG: $VERSION". Open "View more details on Google Cloud Build", and rerun from there.

    Retry release-on-tag from the Google Cloud Build details page

    Verify that you're retrying the right build

    TODO: we should have an automation KFP cluster, and the waiting and submiting release-on-tag cloudbuild task should happen automatically.

    NOTE: postsubmit tests will most likely fail for the release commit, this is expected, postsubmit tests start right after the commit is in GitHub repo, but some artifacts they depend on are still being built by the processes in these two steps.

  7. Search "PyPI" in Google internal release doc for getting password of kubeflow-pipelines user.

  8. Release kfp-server-api python packages to PyPI.

    git checkout $BRANCH
git pull upstream $BRANCH
cd backend/api/v2beta1/python_http_client
rm -r dist
python3 setup.py --quiet sdist
python3 -m twine upload dist/*

  9. Release kfp python packages to PyPI. (Note: Please skip this step for backend release, this step will be handled by SDK release.)

    pip3 install twine --user
gsutil cp gs://ml-pipeline/release/$VERSION/kfp.tar.gz kfp-$VERSION.tar.gz
python3 -m twine upload kfp-$VERSION.tar.gz

    !!! The file name must contain the version. See #1292

  10. Create a GitHub release using $VERSION git tag and title Version $VERSION, fill in the description. Detailed steps:

    1. Draft a new release.

    2. Typing in version tag field to search and select the "$VERSION" tag published in release instructions above. Its format is like X.Y.Z or X.Y.Z-rc.N.

    3. Use this template for public releases and replace the $VERSION with real values.

       To deploy Kubeflow Pipelines in an existing cluster, follow the instruction in [here](https://www.kubeflow.org/docs/pipelines/standalone-deployment-gcp/) or via UI [here](https://console.cloud.google.com/ai-platform/pipelines)

 Install python SDK (python 3.9 above) by running:

 ```bash
 python3 -m pip install kfp kfp-server-api --upgrade
 ```

 See the [Change Log](https://github.com/kubeflow/pipelines/blob/$VERSION/CHANGELOG.md)

      Use this template for prereleases (release candidates) and PLEASE CHECK the This is a prerelease checkbox in the GitHub release UI.

        To deploy Kubeflow Pipelines in an existing cluster, follow the instruction in [here](https://www.kubeflow.org/docs/pipelines/standalone-deployment-gcp/).

  Install kfp-server-api package (python 3.9 above) by running:

  ```bash
  python3 -m pip install kfp-server-api==$VERSION --upgrade
  ```

  Refer to:
  * [Upgrade Notes with notices and breaking changes](https://www.kubeflow.org/docs/components/pipelines/installation/upgrade/)
  * [Change Log](https://github.com/kubeflow/pipelines/blob/$VERSION/CHANGELOG.md)

  NOTE, kfp python SDK is **NOT** included and released separately.

  11. (Do this step only when releasing from a NON-master release branch) Update master branch to the same version and include latest changelog:

    git checkout master
git pull
git checkout -b <your-branch-name>
# This avoids line break at end of line.
echo -n $VERSION > VERSION
# This takes a while.
pushd test/release
make release-in-place
popd
git checkout $VERSION -- CHANGELOG.md
git add -A
git commit -m "chore(release): bump version to $VERSION on master branch"

  12. If current release is not a prerelease, create a PR to update version in kubeflow documentation website: https://github.com/kubeflow/website/blob/master/layouts/shortcodes/pipelines/latest-version.html

    Note, there MUST NOT be a line ending in the file. Editing on GitHub always add a line ending for you so you cannot create a PR on GitHub UI. Instead, you can checkout the repo locally and

    echo -n 1.0.0 > layouts/shortcodes/pipelines/latest-version.html

    and create a PR to update the version, e.g. kubeflow/website#1942.

Release Process Development

Please refer to ./test/release.