Continuous Integration ¶
Using kOps' declarative manifests it is possible to create and manage clusters entirely in a CI environment.
Rather than using
kops create cluster and
kops edit cluster, the cluster and instance group manifests can be stored in version control.
This allows cluster changes to be made through reviewable commits rather than on a local workstation.
This is ideal for larger teams in order to avoid possible collisions from simultaneous changes being made.
It also provides an audit trail, consistent environment, and centralized view for any kOps commands being ran.
Running kOps in a CI environment can also be useful for upgrading kOps. Simply download a newer version in the CI environment and run a new pipeline. This will allow any changes to be reviewed prior to being applied. This strategy can be extended to sequentially upgrade kOps on multiple clusters, allowing changes to be tested on non-production environments first.
This page provides examples for managing kOps clusters in CI environments. The Manifest documentation describes how to create the YAML manifest files locally and includes high level examples of commands described below.
If you have a solution for a different CI platform or deployment strategy, feel free to open a Pull Request!
GitLab CI ¶
GitLab CI is built into GitLab and allows commits to trigger CI pipelines.
- The GitLab runners that run the jobs need the appropriate permissions to invoke the kOps commands. For clusters in AWS this means providing AWS IAM credentials either with IAM User Keys set as secret variables in the project, or having the runner run on an EC2 instance with an instance profile attached.
Example Workflow ¶
- A cluster administrator makes a change to a cluster manifest, commits and pushes to a feature branch on GitLab and opens a Merge Request
- A reviewer reviews the change to confirm it is as intended, and approves or merges the MR
- A "master" pipeline is triggered from this merge commit which runs a
kops update cluster.
- The administrator reviews the output of the
dryrunjob to confirm the desired changes and initiates the
updatejob which runs
kops update cluster --yes.
- Once the cluster is updated,
kops rolling-update clusteris ran which can be used to confirm any nodes that need replacement. The administrator then starts the
rolljob which runs
kops rolling-update cluster --yesand replaces any nodes as necessary.
- kops replace --force -f cluster.yml
- kops update cluster
- kops update cluster --yes
- kops rolling-update cluster
- kops rolling-update cluster --yes
only:field in each job will need to be updated to reflect the real project's namespace and name. The two variables will also need to be set to real values.
- The jobs that make actual changes to the clusters are manually invoked (
when: manual) though this could easily be removed to make them automatic.
- This pipeline setup will create and update existing clusters in place. It does not perform a "blue/green" deployment of multiple clusters.
- The pipeline can be extended to support multiple clusters by making separate jobs per cluster for each stage.
KOPS_CLUSTER_NAMEvariable is set correctly for each set of jobs.
In this case, it is possible to use
kops toolbox template to manage one YAML template and per-cluster values files with which to render the template.
See the Cluster Template documentation for more information.
kops toolbox template would then be ran before
- This pipeline does not have a true "dryrun" job that can be ran on non-master branches, for example before a merge request is merged.
This is because the required
kops replacebefore the
kops update clusterwill update the live assets in the state store which could impact newly launched nodes that download these assets. PR #6465 added support for copying the state store to a local filesystem prior to
kops replace, allowing the dryrun pipeline to be completely isolated from the live state store. For example, run
aws s3 sync "$KOPS_STATE_STORE"/"$CLUSTER_NAME" /some/local/path/"$CLUSTER_NAME" --exclude '*backups/*' --quietfirst, then pass
kops update cluster.