Configuring high-availability components

Active/passive high availability (HA) is a standard feature of Kubernetes APIs that helps to ensure that APIs stay operational if a disruption occurs. In an HA deployment, if an active controller crashes or is deleted, another controller is available to take over processing of the APIs that were being serviced by the controller that is now unavailable.

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.

Disabling leader election

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

Scaling the control plane

With the exception of the activator component you can scale up any deployment running in knative-serving (or kourier-system) with a command like:

  1. $ kubectl -n knative-serving scale deployment <deployment-name> --replicas=2
  • Setting --replicas to a value of 2 enables HA.
  • 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.
  • Setting --replicas=1 disables HA.

NOTE: If you scale down the autoscaler component, 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.

Scaling the data plane

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:

  1. $ kubectl get hpa activator -n knative-serving

The possible output will be something like:

  1. NAME REFERENCE TARGETS MINPODS MAXPODS REPLICAS AGE
  2. activator Deployment/activator 2%/100% 5 15 11 346d

By default minReplicas and maxReplicas 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:

  1. $ kubectl patch hpa activator -n knative-serving -p '{"spec":{"minReplicas":9,"maxReplicas":19}}'

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.