Onboarding a New Component for Testing and Merge Automation

Overview

This document overviews the workflow for onboarding new public component repositories to the Openshift CI. Private repositories are not supported at this time, but will be in the future

Granting Robots Privileges and Installing the GitHub App

In order to add labels, move PRs and issues into milestones, merge PRs, etc, the robots will need write access to your repository.

If your repository is in the OpenShift organization and it was created by Dev Productivity team, it probably has Team OpenShift Robots added by default as part of the repo creation process. To check if your repo has the robots added, please ask in slack in #forum-dp-platform or open a ticket. If your repo does not have robots, then open a new ticket in their JIRA project with the details.

If your component repository is not in this organization:

1. Invite openshift-ci-robot and openshift-merge-robot into your organization or add them as collaborators for the repository.
2. The invitations will be accepted automatically via the periodic-openshift-release-master-accept-invitations-ci-robot and periodic-openshift-release-master-accept-invitations-merge-robot jobs in no more than 4 hours.
3. The openshift-merge-robot must be given admin permissions to support either of the following use cases:
   1. In order for tide to automatically merge PRs it must be allowed to verify the repo’s allowed merge methods.
   2. By default, we enable branch-protection for all prow-controlled repos. We can disable it in prow’s config.yaml
4. If a repository is enrolled in centralized branch management and no write permissions is granted to openshift-merge-robot, ensure that the tide/merge-blocker label exists on the repository. Otherwise, the periodic-openshift-release-merge-blockers job would fail. See how to create a label at Github’s documentation.
5. All repositories need the Openshift CI GitHub App installed. Go to the app and click Configure. We plan to eventually replace the bot accounts entirely with that app, but that work is not yet done.

Prow Configuration

Prow is the k8s-native upstream CI system, source code hosted in the kubernetes/test-infra repository. Prow interacts with GitHub to provide the automation UX that developers use on their pull requests, as well as orchestrating test workloads for those pull requests.

Bootstrapping Configuration for a new Repository

From the root of the openshift/release repository, run the following target and use the interactive tool to bootstrap a configuration for your repository:

1

make new-repo

This should fully configure your repository, so the changes that it produces are ready to be submitted in a pull request. The resulting YAML file called $org-$repo-$branch.yaml will be found in the ci-operator/config/$org/$repo directory.

Enabling Plugins

Plugins implement many portions of CI system functionality. These must be configured before they can be used. They can be enabled as-needed and configured for your repository independently of other repositories in the organization. After your repository is configured to deliver webhooks to Prow, those webhooks will be ignored until plugins are configured to consume and react to them. A live list of plugins and their descriptions is hosted on our website, please consult that list while reading the following section for more detail.

After initializing your repository with the make new-repo target, you can create a new Plugin configuration for your repository with the following target:

1

make prow-config

This will place a new _pluginconfig.yaml file in the /core-services/prow/02_config/$org/$repo directory. This file is used to configure the specific plugins for your repository.

Default plugin configuration is stored in _plugins.yaml in the openshift/release repository. Plugins are enabled for a repository or organization under the plugins key. Plugin-specific configuration is under keys like label or owners. The set of a repository’s enabled plugins is the union of plugins configured for the repository’s organization (found at the plugins.yaml["plugins"]["$org"] key) and the repository itself (found in the /core-services/prow/02_config/$org/$repo/_pluginconfig.yaml file).

Most individual plugins can be configured to change their behavior; only some plugins allow for granular configuration at a repository level, many only expose global configuration options for all repositories that Prow monitors. If you think you need to configure an individual plugin, consult a CI administrator. While we work on a better solution, documentation for these options lives in the type Configuration struct here.

Repositories Under Existing Organizations

If you are onboarding a repository in an organization for which plugins are already configured, you will only need to enable plugins that you do not already inherit from the organization by adding those plugins under a new plugins.yaml["plugins"]["$org/$repo"] key. The one plugin you may want to configure at a repository-scoped level is:

Review the list of plugins enabled for your owning organization and the live plugin catalog when choosing which plugins you want on your repo.

If your repository does not have OWNERS files, you will not be able to opt into the /approve process or automatic pull request review assignment. OWNERS file format and interaction details can be found upstream. You will also need to add bugzilla component information to your OWNERS file. Format can be found here.

Repositories Under New Organizations

If you are onboarding a component not in any organization we already have configured, consider copying the openshift organization’s plugin configuration for your organization under a new plugins.yaml["plugins"]["$org"] key, or/ setting your repo’s configuration without adding any organization-wide configuration by adding the full list of plugins you require under a new plugins.yaml["plugins"]["$org/$repo"] key.

Describing Tests

Prow provides the following test trigger types:

Configuration for your repository’s tests live in YAML files in the openshift/release repository. Jobs are stored in many files, sharded by branch and job type, found under: ci-operator/jobs/$org/$repo/$org-$repo-$branch-$jobtype.yaml. The org and repo name redundancy is because of a requirement that the basename of your YAML file is unique under the ci-operator/jobs tree. More detailed Prow job configuration documentation lives upstream.

Generating Prow jobs from ci-operator configuration files

The Test Platform team created a tool to generate Prow job configuration files out of ci-operator configuration files. The generator has knowledge of the naming and directory structure conventions in openshift/release repository. Provided you have put the ci-operator configuration file to ci-operator/config/$org/$repo directory in it (as described by Containerized Tests section), you can generate the needed Prow files by running this command from the root of the openshift/release repository:

1

make jobs

This will create all necessary files under ci-operator/jobs/$org/$repo, creating a good default set of Prow jobs.

1
2
3

make jobs WHAT=openshift
# or
make jobs WHAT=openshift/release

Setting up team ownership of ci-operator and Prow config files

While the initial PR to openshift/release will need to be reviewed and approved by root approvers, once the component config is in place, it should be owned by the component team. To achieve this, an OWNERS file mirroring what exists upstream should be placed into both ci-operator/config/$org/$repo and ci-operator/jobs/$org/$repo directories.

Enabling Automatic Merges

Prow’s tide component periodically searches for pull requests that fit merge criteria (for instance, presence of a lgtm label and absence of the do-not-merge/hold label) and merges them. Tide furthermore requires not only that all required tests in the Prow configuration succeed and all posted statuses on the GitHub pull request are green but also that the tests tested the latest commit in the pull request on top of the latest commit in the branch that the pull request is targeting before a pull request is considered for merging.

To enable Tide, place a new _prowconfig.yaml file in the /core-services/prow/02_config/$org/$repo directory and configure Tide under the top-level ["tide"] key. The easiest way to get the correct config is to copy it from an existing repo with similar requirements. Tide’s configuration options are documented upstream.

If your repository does not have OWNERS files, or if you have not chosen to opt into the /approve process, it is suggested that you require only the lgtm label and not approve also.

Who can /approve?

Repo’s OWNERS and OWNERS_ALIASES define the list. It is a concept of Prow workflow. Those files also define who could get selected as reviewers of PRs in that repo. See Prow’s doc on this topic.

Who can /lgtm?

Github’s users who are the repo’s collaborators. It is a concept of github. Contributors should follow this mojo to become a collaborator for repositories in openshift org. See Prow’s doc on this topic.

CI Operator Configuration

CI-Operator is a second-level orchestrator which translates Prow’s testing requests into OpenShift-native test workloads. Think translating “run integration tests on my PR” into “trigger an OpenShift Build to create a container with test artifacts and a Pod to run the integration test using that container image”.

While the Prow configuration describes when to run a test, the CI Operator configuration describes the test’s content.

Consult the ci-operator configuration reference document for information on specific fields.

Containerized Tests

Adding a containerized test is as simple as adding an entry to the tests array in the CI Operator configuration file and a Prow job configuration that runs the test with --target=$target. Consult the documentation for more details on how to configure containerized tests and test them locally.

We recommend breaking up your tests into logical sections when adding tests here. More granular test reporting will allow for higher parallelism during test execution and more efficient re-testing if one suite fails.

Containerized tests are configured first-class in CI Operator configuration files in the openshift/release repository, sharded by branch, at: ci-operator/config/$org/$repo/$org-$repo-$branch.yaml. The org and repo name redundancy is due to a requirement that all filenames be unique under ci-operator/config.

End-to-End Tests

Tests that require a running Openshift cluster should use one of the provided step registry workflows, more details here.

Image Publishing and Mirroring

When container images are declared as release artifacts for a repository in the CI Operator configuration file under the images list (like this), a synthetic [images] target is available for CI Operator execution that will simply build all release images. In order for the container images built from your repository to be published, the Prow job generator will configure a Prow postsubmit job that uses the CI Operator --target=[images] and --promote flags. Information on how to publish images to an external registry can be found in a separate document..

Product builds and becoming part of an OpenShift release image

Some images become part of the OpenShift release image because they are core to the platform. To be part of the release image, there are additional requirements (beyond those described in the previous section).

There are two types of component images in the release payloads:

1. Operators managed by the CVO - known as second level operators. If a Dockerfile contains LABEL io.openshift.release.operator true, the component is a second level operator.
2. Operands managed by second level operators. These images are pulled into the release payload by virtue of being specified in an image-references file by a second level operator.

All product teams that will ship an image to product must ensure their image is built in OSBS at least once and published back to the nightly test jobs BEFORE you reference them from another component (via the image-references file), or before you set the image label io.openshift.release.operator to get automatically included.

1. To include the build version, you can consume environment variables like OS_GIT_VERSION and BUILD_VERSION set by ART’s Doozer in your Dockerfile. For example, here is build-machinery-go passing OS_GIT_VERSION, SOURCE_GIT_COMMIT, and SOURCE_GIT_TREE_STATE through to Go.

   CVO manifests may also use the 0.0.1-snapshot placeholder to have the OCP release version injected at release-assembly time.

2. Ensure you have successfully published your image to the CI integration stream

3. Follow the ART instructions to have them build your image

   • On the dist-git part of the process it is critical that you ensure your component/image names match as described in the bulleted criteria

4. Ensure a single successful build is run (sync with ART to confirm)

5. Open the PR to add LABEL io.openshift.release.operator true or to add your new image to another component (your operator, usually)

6. Once the PR is merged, verify that the nightly builds continue to pass (usually 2-3 hours after your PR merges) and that you didn’t break the OCP CI test

Renaming or removing components in the OpenShift release payload

It is occasionally necessary to rename or remove existing components in the OpenShift release payload. This must be done with care. Missteps may cause production and CI automation to be unable to create new release payloads – impacting the entire organization. Observing the process as it unfolds and maintaining good communication with the ART team is crucial.

Changing the component name of a second level operator

It is uncommon for other components to directly reference the release payload component name of a second level operator. However, if your second level operator is referenced, follow the procedure for changing the component name of an operand.

1
2
3

$ oc adm release new --from-release {latest-4.x-nightly-pullspec} --exclude etcd --to-file output.yaml
.....
error: unable to create a release: operator "cluster-etcd-operator" contained an invalid image-references file: no input image tag named "etcd"

Failure to follow all steps can lead to important and difficult to detect disparities between what is tested in CI and what is shipped to customers.

Steps:

1. If a name change was not directly requested by a staff engineer, ensure that a staff engineer agrees on the name change (@aos-staff-engineers on Slack).
2. Open and /hold a pull request, PR1, against github.com/openshift/release to change the component’s name in CI. In virtually all cases, this will be a PR against the component’s ci-config for its main branch. A component’s name should be specified in the images.to: stanza of a component’s CI configuration. The new tag name should match exactly what will appear in oc adm release info when the component is listed.
3. Open and /hold a pull request, PR2, against github.com/openshift/ocp-build-data in the openshift-4.x branch of the targeted release(s). Note that ART may have already cut a new release branch, meaning you need to open a PR for openshift-4.x+1 as well. The PR should change the name field in the component’s metadata. Note that in ART metadata, the desired name should be prefixed with ose-. For example, the cluster-etcd-operator payload component name is defined here, for openshift-4.11.
4. Submit a copy of this ART team template to communicate to ART that a component name change is desired. Include the PRs in the Jira ticket.
5. All PRs should be passing tests and ready to merge. No PRs other than PR1 should be merged in the component’s github.com repository during the following process. In a synchronous chat with an ART team member or release manager (@release-artists in #aos-art on Slack) the following should be performed:
   1. On the central app.ci CI cluster, a release-artist should check the current registry.ci.openshift.org image associated with the component in the -n ocp is/4.{minor} release image stream. For example, oc -n ocp get istag 4.11:{old-component-name} -o=json | jq .tag.from.name should output a pullspec like registry.ci.openshift.org/ocp/4.11@sha256:... for 4.11.
   2. Unhold PR1 and allow it to merge.
   3. Once PR1 merges, the release-artist should run oc -n ocp tag {existing-registry.ci-openshift-org@sha26..} 4.{minor}:{new-component-name} followed immediately by oc -n ocp tag 4.{minor}:{old-component-name} -d to remove the old component name from CI.
   4. Unhold PR2 and have the release-artist merge it.
6. Monitor the subsequent CI payloads on amd64 OpenShift release controller. Continue to do so until you see a CI payload produced which reports the new component name in oc adm release info <ci-release-payload-pullspec>. If steps in this process were missed, following the hyperlink for a payload name in the release controller will simply display an error message stating that the release controller was unable to create a release with a few details about the problem’s cause. If this error is displayed, immediately report the issue to @team-technical-release and @release-artists so that the incident can be recovered as quickly as possible.

Changing the component name of an operand

Release payload operand component names are referenced in second level operator github.com repositories. A change must merge in the second level operator’s image-references file, or release payloads will fail to assemble after the operand component name is changed.

Steps:

1. If a name change was not directly requested by a staff engineer, ensure that a staff engineer agrees on the name change (@aos-staff-engineers on Slack).
2. Open and /hold a pull request, PR1, against github.com/openshift/release to change the operand component’s name in CI. In virtually all cases, this will be a PR against the component’s ci-config for its main branch. A component’s name should be specified in the images.to: stanza of a component’s CI configuration. The new tag name should match exactly what will appear in oc adm release info when the component is listed.
3. Open and /hold a pull request, PR2, against the repo(s) of the operator(s) which references the old operand component name in its image-references file. PR2 should update image-references to use the new component name.
4. Open and /hold a pull request, PR3, against github.com/openshift/ocp-build-data in the openshift-4.x branch of the targeted release(s). Note that ART may have already cut a new release branch, meaning you need to open a PR for openshift-4.x+1 as well. The PR should change the name field in the component’s metadata. Note that in ART metadata, the desired name should be prefixed with ose-. For example, the etcd operand component is name definition can be seen here, for openshift-4.11.
5. Submit a copy of this ART team template to communicate to ART that a component name change is desired. Include the PRs in the Jira ticket.
6. PR1 and PR3 should be passing tests and ready to merge (PR2 will be failing at this point). No PRs other than PR1 and PR2 should be merged in their respective github.com repositories during the following process. In a synchronous chat with an ART team member or release manager (@release-artists in #aos-art on Slack) the following should be performed:
   1. On the central app.ci CI cluster, a release-artist should check the current registry.ci.openshift.org image associated with the component in the -n ocp is/4.{minor} release image stream. For example, oc -n ocp get istag 4.11:{old-component-name} -o=json | jq .tag.from.name should output a pullspec like registry.ci.openshift.org/ocp/4.11@sha256:... for 4.11.
   2. The release-artist should run oc -n ocp tag {existing-registry.ci-openshift-org@sha26..} 4.{minor}:{new-component-name} to establish a tag with the new operand component name for CI.
   3. Run /retest on PR2(s). Tests should now pass.
   4. Unhold and merge PR1.
   5. Unhold and merge PR2(s).
   6. Unhold and have the release-artist merge PR3. The time between PR2 and PR3 merging should be kept to a minimum to avoid ART nightlies failing to assemble.
   7. It is not time sensitive, but before the ART Jira ticket is closed, the release-artist must remove the old component name from CI: oc -n ocp tag 4.{minor}:{old-component-name} -d.
7. Monitor the subsequent CI payloads on amd64 OpenShift release controller. Continue to do so until you see a CI payload produced which reports the new component name in oc adm release info <ci-release-payload-pullspec>. If steps in this process were missed, following the hyperlink for a payload name in the release controller will simply display an error message stating that the release controller was unable to create a release with a few details about the problem’s cause. If this error is displayed, immediately report the issue to @team-technical-release and @release-artists so that the incident can be recovered as quickly as possible.
8. During the ensuing work day, check ART nightlies on the s390x release controller. Continue to do so until you see an ART nightly produced which reports the new component name in oc adm release info <art-s390x-nightly-release-payload-pullspec>. If steps in this process were missed, following the hyperlink for a payload name in the release controller will simply display an error message stating that the release controller was unable to create a release with a few details about the problem’s cause. If this error is displayed, immediately report the issue to @team-technical-release and @release-artists so that the incident can be recovered as quickly as possible. Changes will also eventually be apparent on the amd64 release controller, but, due to differences in acceptance testing, they will be evident in the s390x release controller much sooner.

Removing a component from the OpenShift release payload

Steps:

1. Submit a copy of this ART team template to communicate to ART that a component change is desired. Take care to mention whether the image should also be removed from the 4.x+1 branch of ART’s metadata in case it has already been branched.
2. Open and /hold pull request, PR1, removing references to the old component from any second level operator which includes the component in its image-references file.
3. Open and /hold a pull request, PR2, to github.com/openshift/release which, minimally, removes the promotion: stanza from the ci-operator configuration for the component & affected release(s). If the component is being completely removed from CI, PR2 can be the complete deletion of the ci-operator configuration file for the component / branch.
4. Via the ART Jira ticket, have a release-artist prepare a github.com/openshift/ocp-build-data pull request, PR3, which either removes the component metadata or prevents its inclusion in the release payload (for_payload: false).
5. In a synchronous chat with an ART team member or release manager (@release-artists in #aos-art on Slack) the following should be performed:
   1. Unhold and merge PR1.
   2. Unhold and merge PR2.
   3. Unhold and have the release-artist merge PR3.
   4. The release-artist should then remove the old component name from CI: oc -n ocp tag 4.{minor}:{old-component-name} -d.