Ephemeral Metadata

Important

Available for canary rollouts since v0.10.0

Important

Available for blue-green rollouts since v1.0

One use case is for a Rollout to label or annotate the desired/stable pods with user-defined labels/annotations, for only the duration which they are the desired or stable set, and for the labels to be updated/removed as soon as the ReplicaSet switches roles (e.g. from desired to stable). The use case which this enables, is to allow prometheus, wavefront, datadog queries and dashboards to be built, which can rely on a consistent labels, rather than the rollouts-pod-template-hash which is unpredictable and changing from revision to revision.

A Rollout using the canary strategy has the ability to attach ephemeral metadata to the stable or canary Pods using the stableMetadata and canaryMetadata fields respectively.

spec:
  strategy:
    canary:
      stableMetadata:
        labels:
          role: stable
      canaryMetadata:
        labels:
          role: canary

A Rollout using the blue-green strategy has the ability to attach ephemeral metadata to the active or preview Pods using the activeMetadata and previewMetadata fields respectively.

spec:
  strategy:
    blueGreen:
      activeMetadata:
        labels:
          role: active
      previewMetadata:
        labels:
          role: preview

During an update, the Rollout will create the desired ReplicaSet while also merging the metadata defined in canaryMetadata/previewMetadata to the desired ReplicaSet's spec.template.metadata. This results in all Pods of the ReplicaSet being created with the desired metadata. When the rollout becomes fully promoted, the desired ReplicaSet becomes the stable, and is updated to use the labels and annotations under stableMetadata/activeMetadata. The Pods of the ReplicaSet will then be updated in place to use the stable metadata (without recreating the pods).

Important

In order for tooling to take advantage of this feature, they would need to recognize the change in labels and/or annotations that happen after the Pod has already started. Not all tools may detect this.