Chapter 2. Autoscaling

2.1. Autoscaling

Knative Serving provides automatic scaling, or autoscaling, for applications to match incoming demand. For example, if an application is receiving no traffic, and scale-to-zero is enabled, Knative Serving scales the application down to zero replicas. If scale-to-zero is disabled, the application is scaled down to the minimum number of replicas configured for applications on the cluster. Replicas can also be scaled up to meet demand if traffic to the application increases.

Autoscaling settings for Knative services can be global settings that are configured by cluster administrators (or dedicated administrators for Red Hat OpenShift Service on AWS and OpenShift Dedicated), or per-revision settings that are configured for individual services.

You can modify per-revision settings for your services by using the OpenShift Container Platform web console, by modifying the YAML file for your service, or by using the Knative (kn) CLI.

Note

Any limits or targets that you set for a service are measured against a single instance of your application. For example, setting the target annotation to 50 configures the autoscaler to scale the application so that each revision handles 50 requests at a time.

2.2. Scale bounds

Scale bounds determine the minimum and maximum numbers of replicas that can serve an application at any given time. You can set scale bounds for an application to help prevent cold starts or control computing costs.

2.2.1. Minimum scale bounds

The minimum number of replicas that can serve an application is determined by the min-scale annotation. If scale to zero is not enabled, the min-scale value defaults to 1.

The min-scale value defaults to 0 replicas if the following conditions are met:

  • The min-scale annotation is not set
  • Scaling to zero is enabled
  • The class KPA is used

Example service spec with min-scale annotation

apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: showcase
  namespace: default
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/min-scale: "0"
...

2.2.1.1. Setting the min-scale annotation by using the Knative CLI

Using the Knative (kn) CLI to set the min-scale annotation provides a more streamlined and intuitive user interface over modifying YAML files directly. You can use the kn service command with the --scale-min flag to create or modify the min-scale value for a service.

Prerequisites

  • Knative Serving is installed on the cluster.
  • You have installed the Knative (kn) CLI.

Procedure

  • Set the minimum number of replicas for the service by using the --scale-min flag: 

    $ kn service create <service_name> --image <image_uri> --scale-min <integer>

    Example command

    $ kn service create showcase --image quay.io/openshift-knative/showcase --scale-min 2

2.2.2. Maximum scale bounds

The maximum number of replicas that can serve an application is determined by the max-scale annotation. If the max-scale annotation is not set, there is no upper limit for the number of replicas created.

Example service spec with max-scale annotation

apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: showcase
  namespace: default
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/max-scale: "10"
...

2.2.2.1. Setting the max-scale annotation by using the Knative CLI

Using the Knative (kn) CLI to set the max-scale annotation provides a more streamlined and intuitive user interface over modifying YAML files directly. You can use the kn service command with the --scale-max flag to create or modify the max-scale value for a service.

Prerequisites

  • Knative Serving is installed on the cluster.
  • You have installed the Knative (kn) CLI.

Procedure

  • Set the maximum number of replicas for the service by using the --scale-max flag: 

    $ kn service create <service_name> --image <image_uri> --scale-max <integer>

    Example command

    $ kn service create showcase --image quay.io/openshift-knative/showcase --scale-max 10

2.3. Concurrency

Concurrency determines the number of simultaneous requests that can be processed by each replica of an application at any given time. Concurrency can be configured as a soft limit or a hard limit:

  • A soft limit is a targeted requests limit, rather than a strictly enforced bound. For example, if there is a sudden burst of traffic, the soft limit target can be exceeded.

  • A hard limit is a strictly enforced upper bound requests limit. If concurrency reaches the hard limit, surplus requests are buffered and must wait until there is enough free capacity to execute the requests.

    Important

    Using a hard limit configuration is only recommended if there is a clear use case for it with your application. Having a low, hard limit specified may have a negative impact on the throughput and latency of an application, and might cause cold starts.

Adding a soft target and a hard limit means that the autoscaler targets the soft target number of concurrent requests, but imposes a hard limit of the hard limit value for the maximum number of requests.

If the hard limit value is less than the soft limit value, the soft limit value is tuned down, because there is no need to target more requests than the number that can actually be handled.

2.3.1. Configuring a soft concurrency target

A soft limit is a targeted requests limit, rather than a strictly enforced bound. For example, if there is a sudden burst of traffic, the soft limit target can be exceeded. You can specify a soft concurrency target for your Knative service by setting the autoscaling.knative.dev/target annotation in the spec, or by using the kn service command with the correct flags.

Procedure

  • Optional: Set the autoscaling.knative.dev/target annotation for your Knative service in the spec of the Service custom resource:

    Example service spec

    apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: showcase
  namespace: default
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/target: "200"

  • Optional: Use the kn service command to specify the --concurrency-target flag: 

    $ kn service create <service_name> --image <image_uri> --concurrency-target <integer>

    Example command to create a service with a concurrency target of 50 requests

    $ kn service create showcase --image quay.io/openshift-knative/showcase --concurrency-target 50

2.3.2. Configuring a hard concurrency limit

A hard concurrency limit is a strictly enforced upper bound requests limit. If concurrency reaches the hard limit, surplus requests are buffered and must wait until there is enough free capacity to execute the requests. You can specify a hard concurrency limit for your Knative service by modifying the containerConcurrency spec, or by using the kn service command with the correct flags.

Procedure

  • Optional: Set the containerConcurrency spec for your Knative service in the spec of the Service custom resource:

    Example service spec

    apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: showcase
  namespace: default
spec:
  template:
    spec:
      containerConcurrency: 50

    The default value is 0, which means that there is no limit on the number of simultaneous requests that are permitted to flow into one replica of the service at a time.

    A value greater than 0 specifies the exact number of requests that are permitted to flow into one replica of the service at a time. This example would enable a hard concurrency limit of 50 requests.

  • Optional: Use the kn service command to specify the --concurrency-limit flag: 

    $ kn service create <service_name> --image <image_uri> --concurrency-limit <integer>

    Example command to create a service with a concurrency limit of 50 requests

    $ kn service create showcase --image quay.io/openshift-knative/showcase --concurrency-limit 50

2.3.3. Concurrency target utilization

This value specifies the percentage of the concurrency limit that is actually targeted by the autoscaler. This is also known as specifying the hotness at which a replica runs, which enables the autoscaler to scale up before the defined hard limit is reached.

For example, if the containerConcurrency value is set to 10, and the target-utilization-percentage value is set to 70 percent, the autoscaler creates a new replica when the average number of concurrent requests across all existing replicas reaches 7. Requests numbered 7 to 10 are still sent to the existing replicas, but additional replicas are started in anticipation of being required after the containerConcurrency value is reached.

Example service configured using the target-utilization-percentage annotation

apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: showcase
  namespace: default
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/target-utilization-percentage: "70"
...

2.4. Scale-to-zero

Knative Serving provides automatic scaling, or autoscaling, for applications to match incoming demand.

2.4.1. Enabling scale-to-zero

You can use the enable-scale-to-zero spec to enable or disable scale-to-zero globally for applications on the cluster.

Prerequisites

  • You have installed OpenShift Serverless Operator and Knative Serving on your cluster.
  • You have cluster administrator permissions on OpenShift Container Platform, or you have cluster or dedicated administrator permissions on Red Hat OpenShift Service on AWS or OpenShift Dedicated.
  • You are using the default Knative Pod Autoscaler. The scale to zero feature is not available if you are using the Kubernetes Horizontal Pod Autoscaler.

Procedure

  • Modify the enable-scale-to-zero spec in the KnativeServing custom resource (CR):

    Example KnativeServing CR

    apiVersion: operator.knative.dev/v1beta1
kind: KnativeServing
metadata:
  name: knative-serving
spec:
  config:
    autoscaler:
      enable-scale-to-zero: "false" 1

    1
    The enable-scale-to-zero spec can be either "true" or "false". If set to true, scale-to-zero is enabled. If set to false, applications are scaled down to the configured minimum scale bound. The default value is "true".

2.4.2. Configuring the scale-to-zero grace period

Knative Serving provides automatic scaling down to zero pods for applications. You can use the scale-to-zero-grace-period spec to define an upper bound time limit that Knative waits for scale-to-zero machinery to be in place before the last replica of an application is removed.

Prerequisites

  • You have installed OpenShift Serverless Operator and Knative Serving on your cluster.
  • You have cluster administrator permissions on OpenShift Container Platform, or you have cluster or dedicated administrator permissions on Red Hat OpenShift Service on AWS or OpenShift Dedicated.
  • You are using the default Knative Pod Autoscaler. The scale-to-zero feature is not available if you are using the Kubernetes Horizontal Pod Autoscaler.

Procedure

  • Modify the scale-to-zero-grace-period spec in the KnativeServing custom resource (CR):

    Example KnativeServing CR

    apiVersion: operator.knative.dev/v1beta1
kind: KnativeServing
metadata:
  name: knative-serving
spec:
  config:
    autoscaler:
      scale-to-zero-grace-period: "30s" 1

    1
    The grace period time in seconds. The default value is 30 seconds.