Upgrade server

Server v4.3

Server Admin

Upgrade paths

We recommend that you do not skip releases when upgrading, however some patch releases can be skipped. Your upgrade path will depend on your current version, and the currently available version releases.

When upgrading from one minor version to the next, ensure your upgrade path includes the last patch release for your current minor version.

To see some common upgrade path options, see this support article.

Notes on changes in v4.3

Please note the following significant changes in server v4.3.

Docker layer caching (DLC)

Old DLC volumes will not be carried over and need to be wiped manually
Projects that stop using DLC will not delete final DLC cache without manual intervention
DLC now runs through S3 and GCS instead of SSD Volumes

Machine jobs

Machine jobs will now run through machine provisioner
When migrating vm-gc will need to clean up all machines left running before it is removed. All other vm-service related pods can be removed during migration first. If this is not followed, vm-service machines will not be cleaned up without manual intervention. This can alternatively be done by manually cleaning up remaining VMs before starting machine provisioner.
Machine and remote Docker jobs no longer flow through Nomad, and they run completely in a single machine. Nomad is not required to run machine jobs.
If deploying in a specific zone need to specify subnetwork for GCP
IAM users have changed names to end with machine-provisioner instead of vm-service so some users might need to be altered.

Docker jobs

Run through Nomad, but are orchestrated through docker-provisioner

Storage

Excluding DLC, historical artifacts will continue to be supported in the new execution system with no user changes

PostgreSQL

The PostgreSQL image will be updated during this release, which will cause a short downtime.

Workflows conductor

There is a workflows-conductor data migration that will run as a job in the background after your upgrade. The migration processes projects in batches of 1000 and sleeps for 60 seconds before starting another. Processing time for a batch depends on the MongoDB and PostgreSQL data stores, but in our production environment, a batch of 1000 projects took around 6 seconds.
Ensure that your CircleCI server v4.3 installation is left running until after the migration has completed. You can confirm that the migration has completed by checking the logs of workflows-conductor-event-consumer. There will be a starting next-build-seq-migration log message with project_count=0

Vault

We have moved away from Vault to Tink for encryption. The process for migration is documented here, and includes a convenience script to move existing secrets. You should complete the migration to Tink on your v4.2.x installation before backing up your server installation in preparation for upgrading to v4.3. Customers that do not perform this step may have issues restoring Vault from backup in v4.3.

Prerequisites

Ensure you have access to the Kubernetes cluster in which server is installed.
Ensure you have migrated from Vault to Tink on your 4.2.x instance as documented here.
Ensure you have set up Backup and Restore.
Ensure there is a recent backup. For more information, see the Backup and Restore guide.

Upgrade steps

This upgrade is a disruptive process and downtime is expected. Please do not attempt to run jobs during this upgrade.

Ensure your cluster is running a compatible kubernetes version for this release (1.26 - 1.27).
Check the changelog and make sure there are no actions you need to take before deploying a new version.
Rename the vm_service block in your values.yml file to machine_provisioner. This block will use the same values as vm_service with the followed exceptions:
- On AWS, tags are no longer formatted as an array and are instead in dictionary format, for example:
  machine_provisioner: providers: ec2: ... tags: tag1: "value1" tag2: "value2"
- On GCP, zones are now entered as an array, for example:
  machine_provisioner: providers: gcp: ... zones: - <gcp-zone1> - <gcp-zone2>
Update vm-service IAM Policy (machine-provisioner requires slightly different permissions). See the Set up authentication page for more details.

Update the vm-service trust policy. The current policy is directed at the vm-service service account. Change this to machine-provisioner as below:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Federated": "<OIDC_PROVIDER_ARN>"
      },
      "Action": "sts:AssumeRoleWithWebIdentity",
      "Condition": {
        "StringEquals": {
          "<OIDC_PROVIDER_URL>:sub": "system:serviceaccount:<K8S_NAMESPACE>:machine-provisioner"
        }
      }
    }

  ]
}

Optionally, confirm what the update is going to do using Helm Diff:

helm diff upgrade circleci-server oci://cciserver.azurecr.io/circleci-server -n $namespace --version <version> -f <path-to-values.yaml> --username $USERNAME --password $PASSWORD

Perform the upgrade:

helm upgrade circleci-server oci://cciserver.azurecr.io/circleci-server -n $namespace --version <version> -f <path-to-values.yaml> --username $USERNAME --password $PASSWORD

Deploy and run reality check in your test environment to ensure your installation is fully operational.
Remove port 2376 from your vm-service security group as it is no longer needed.