v0.6.0

Migration Guide: v0.6.0

1. What Changed

Resource Reservation Namespace

• Old: runai-reservation
• New: kai-resource-reservation

The name of the resource reservation namespace was changed from runai-reservation to kai-resource-reservation. In order to safely migrate from the previous namespace, GPU sharing workloads must be deleted before the upgrade and reservation pods should not be found in runai-reservation namespace.

The following command should return zero pods before you upgrade:

kubectl get pods -n runai-reservation

Scheduling queue label key

• Old: runai/queue
• New: kai.scheduler/queue

The label key for a scheduling queue was changed from runai/queue to kai.scheduler/queue. In order to adopt the new label key, all workloads must have the new label with the name of the respective queue before the upgrade.

The following command should result without any existing pods:

kubectl get pods -A -l 'runai/queue'

Documentation and examples

Docs and examples have been updated to reflect these changes.

2. Impact

• Any existing reservation pods or GPU-sharing workloads still in runai-reservation will not be managed after upgrade.
• Workloads labeled with runai/queue will not be scheduled by the upgraded version of the scheduler.

3. Pin to Legacy Settings

If adopting these changes is not possible at current time, you can keep using the old namespace and label key by overriding the defaults with values.yaml. To do this, add the following flag to the helm upgrade command. This will configure KAI Scheduler to use the old values until the cluster is updated to the new values:

--values ./docs/migrationguides/v0.6.0/values.yaml