Configuring autoscaling

Knative uses a single shared autoscaler. This is, by default, the Knative Pod Autoscaler (KPA), which provides fast, request-based autoscaling capabilities out of the box.

You can also configure Knative to use Horizontal Pod Autoscaler (HPA), or use your own autoscaler, by creating a controller (also referred to as a reconciler) for the Pod Autoscaler custom resource.

Modifying the ConfigMap for KPA

To modify the KPA configuration, you must modify a Kubernetes ConfigMap called config-autoscaler in the knative-serving namespace.

You can view the default contents of this ConfigMap using the following command.

kubectl -n knative-serving describe cm config-autoscaler

Example of the default Kubernetes ConfigMap

apiVersion: v1
kind: ConfigMap
metadata:
 name: config-autoscaler
 namespace: knative-serving
data:
 container-concurrency-target-default: "100"
 container-concurrency-target-percentage: "0.7"
 enable-scale-to-zero: "true"
 max-scale-up-rate: "1000"
 max-scale-down-rate: "2"
 panic-window-percentage: "10"
 panic-threshold-percentage: "200"
 scale-to-zero-grace-period: "30s"
 stable-window: "60s"
 tick-interval: "2s"
 target-burst-capacity: "200"
 requests-per-second-target-default: "200"

Configuring scale to zero for KPA

To correctly configure autoscaling to zero for revisions, you must modify the following parameters in the ConfigMap.

scale-to-zero-grace-period

scale-to-zero-grace-period specifies the time an inactive revision is left running before it is scaled to zero (min: 6s).

scale-to-zero-grace-period: "30s"

stable-window

When operating in a stable mode, the autoscaler operates on the average concurrency over the stable window (min: 6s).

stable-window: "60s"

stable-window can also be configured in the Revision template as an annotation.

autoscaling.knative.dev/window: "60s"

enable-scale-to-zero

Ensure that enable-scale-to-zero is set to true, if scale to zero is desired.

Termination period

The termination period is the time that the pod takes to shut down after the last request is finished. The termination period of the pod is equal to the sum of the values of the stable-window and scale-to-zero-grace-period parameters. In the case of this example, the termination period would be at least 90s.

Configuring concurrency

Concurrency for autoscaling can be configured using the following methods.

Configuring concurrent request limits

target

target defines how many concurrent requests are wanted at a given time (soft limit) and is the recommended configuration for autoscaling in Knative.

The default value for concurrency target is specified in the ConfigMap as 100.

container-concurrency-target-default: "100"

This value can be configured by adding or modifying the autoscaling.knative.dev/target annotation value in the revision template.

autoscaling.knative.dev/target: "50"

containerConcurrency

NOTE: containerConcurrency should only be used if there is a clear need to limit how many requests reach the app at a given time. Using containerConcurrency is only advised if the application needs to have an enforced constraint of concurrency.

containerConcurrency limits the amount of concurrent requests are allowed into the application at a given time (hard limit), and is configured in the revision template.

containerConcurrency: 0 | 1 | 2-N
  • A containerConcurrency value of 1 will guarantee that only one request is handled at a time by a given instance of the revision container, though requests might be queued, waiting to be served.
  • A value of 2 or more will limit request concurrency to that value.
  • A value of 0 means the system should decide.

containerConcurrency takes precedence over the target values.

Configuring scale bounds (minScale and maxScale)

The minScale and maxScale annotations can be used to configure the minimum and maximum number of pods that can serve applications. These annotations can be used to prevent cold starts or to help control computing costs.

minScale and maxScale can be configured as follows in the revision template;

spec:
 template:
  metadata:
   annotations:
    autoscaling.knative.dev/minScale: "2"
    autoscaling.knative.dev/maxScale: "10"

Using these annotations in the revision template will propagate this to PodAutoscaler objects. PodAutoscaler objects are mutable and can be further modified later without modifying anything else in the Knative Serving system.

kubectl edit podautoscaler <revision-name>

NOTE: These annotations apply for the full lifetime of a revision. Even when a revision is not referenced by any route, the minimal pod count specified by minScale will still be provided. Keep in mind that non-routeable revisions may be garbage collected, which enables Knative to reclaim the resources.

Default behavior

If the minScale annotation is not set, pods will scale to zero (or to 1 if enable-scale-to-zero is false per the ConfigMap mentioned above).

If the maxScale annotation is not set, there will be no upper limit for the number of pods created.

Configuring Horizontal Pod Autoscaler (HPA)

NOTE: You can configure Knative autoscaling to work with either the default KPA or a CPU based metric, i.e. Horizontal Pod Autoscaler (HPA).

You can configure Knative to use CPU based autoscaling instead of the default request based metric by adding or modifying the autoscaling.knative.dev/class and autoscaling.knative.dev/metric values as annotations in the revision template.

spec:
 template:
  metadata:
   annotations:
    autoscaling.knative.dev/metric: cpu
    autoscaling.knative.dev/target: "70"
    autoscaling.knative.dev/class: hpa.autoscaling.knative.dev

It is recommended to use the autoscaling-base-reconciler as implemented in Knative Serving.

To use this reconciler, ensure that you are calling ReconcileSKS from the autoscaling-base-reconciler.

If you want to use metrics collected by Knative like concurrency, ensure that you are using ReconcileMetric to enable that system.

Implementing your own Pod Autoscaler

The Pod Autoscaler custom resource allows you to implement your own autoscaler without changing anything else about the Knative Serving system.

You can implement your own Pod Autoscaler if the requirements of your workload cannot be covered by the KPA or HPA, for example if you want to use a more specialized autoscaling algorithm, or if you need to use a specialized set of metrics not supported by Knative out of the box.

To implement your own Pod Autoscaler, you must create a controller to handle your class of Pod Autoscaler.

To do this, you can copy a Knative sample controller and modify its configuration to suit your desired use case.

For example, if your service's template YAML includes a class annotation like:

autoscaling.knative.dev/class: sample

Your controller should only reconcile PodAutoscaler resources with that target.

The informer setup of your controller might look like this:

paInformer.Informer().AddEventHandler(cache.FilteringResourceEventHandler{
  FilterFunc: reconciler.AnnotationFilterFunc(autoscaling.ClassAnnotationKey, "sample", false),
  Handler:    controller.HandleAll(impl.Enqueue),
})

Additional resources