Configuring high-availability components

    When using a leader election HA pattern, instances of controllers are already scheduled and running inside the cluster before they are required. These controller instances compete to use a shared resource, known as the leader election lock. The instance of the controller that has access to the leader election lock resource at any given time is referred to as the leader.

    Leader election is enabled by default for all Knative Serving components. HA functionality is disabled by default for all Knative Serving components, which are configured with only one replica.

    For components leveraging leader election to achieve HA, this capability can be disabled by passing the flag: . This option will go away when HA graduates to “stable”.

    • You can use a higher value if you have a use case that requires more replicas of a deployment. For example, if you require a minimum of 3 controller deployments, set --replicas=3.

    Note

    If you scale down the Autoscaler, you may observe inaccurate autoscaling results for some Revisions for a period of time up to the stable-window value. This is because when an autoscaler pod is terminating, ownership of the revisions belonging to that pod is passed to other autoscaler pods that are on stand by. The autoscaler pods that take over ownership of those revisions use the stable-window time to build the scaling metrics state for those Revisions.

    The scale of the activator component is governed by the Kubernetes HPA component. You can see the current HPA scale limits and the current scale by running:

    By default minReplicas and are set to 1 and 20, correspondingly. If those values are not desirable for some reason, then, for example, you can change those values to minScale=9 and maxScale=19 using the following command:

    To set the activator scale to a particular value, just set minScale and maxScale to the same desired value.

    It is recommended for production deployments to run at least 3 activator instances for redundancy and avoiding single point of failure if a Knative service needs to be scaled from 0.