Custom Metrics Autoscaler Operator overview - Automatically scaling pods with the Custom Metrics Autoscaler Operator | Nodes

As a developer, you can use Custom Metrics Autoscaler Operator for Red Hat OpenShift to specify how OKD should automatically increase or decrease the number of pods for a deployment, stateful set, custom resource, or job based on custom metrics that are not based only on CPU or memory.

The Custom Metrics Autoscaler Operator is an optional operator, based on the Kubernetes Event Driven Autoscaler (KEDA), that allows workloads to be scaled using additional metrics sources other than pod metrics.

The custom metrics autoscaler currently supports only the Prometheus, CPU, memory, and Apache Kafka metrics.

The Custom Metrics Autoscaler Operator scales your pods up and down based on custom, external metrics from specific applications. Your other applications continue to use other scaling methods. You configure triggers, also known as scalers, which are the source of events and metrics that the custom metrics autoscaler uses to determine how to scale. The custom metrics autoscaler uses a metrics API to convert the external metrics to a form that OKD can use. The custom metrics autoscaler creates a horizontal pod autoscaler (HPA) that performs the actual scaling.

To use the custom metrics autoscaler, you create a ScaledObject or ScaledJob object for a workload, which is a custom resource (CR) that defines the scaling metadata. You specify the deployment or job to scale, the source of the metrics to scale on (trigger), and other parameters such as the minimum and maximum replica counts allowed.

You can create only one scaled object or scaled job for each workload that you want to scale. Also, you cannot use a scaled object or scaled job and the horizontal pod autoscaler (HPA) on the same workload.

The custom metrics autoscaler, unlike the HPA, can scale to zero. If you set the minReplicaCount value in the custom metrics autoscaler CR to 0, the custom metrics autoscaler scales the workload down from 1 to 0 replicas to or up from 0 replicas to 1. This is known as the activation phase. After scaling up to 1 replica, the HPA takes control of the scaling. This is known as the scaling phase.

Some triggers allow you to change the number of replicas that are scaled by the cluster metrics autoscaler. In all cases, the parameter to configure the activation phase always uses the same phrase, prefixed with activation. For example, if the threshold parameter configures scaling, activationThreshold would configure activation. Configuring the activation and scaling phases allows you more flexibility with your scaling policies. For example, you can configure a higher activation phase to prevent scaling up or down if the metric is particularly low.

The activation value has more priority than the scaling value in case of different decisions for each. For example, if the threshold is set to 10, and the activationThreshold is 50, if the metric reports 40, the scaler is not active and the pods are scaled to zero even if the HPA requires 4 instances.

Figure 1. Custom metrics autoscaler workflow

You create or modify a scaled object custom resource for a workload on a cluster. The object contains the scaling configuration for that workload. Prior to accepting the new object, the OpenShift API server sends it to the custom metrics autoscaler admission webhooks process to ensure that the object is valid. If validation succeeds, the API server persists the object.
The custom metrics autoscaler controller watches for new or modified scaled objects. When the OpenShift API server notifies the controller of a change, the controller monitors any external trigger sources, also known as data sources, that are specified in the object for changes to the metrics data. One or more scalers request scaling data from the external trigger source. For example, for a Kafka trigger type, the controller uses the Kafka scaler to communicate with a Kafka instance to obtain the data requested by the trigger.
The controller creates a horizontal pod autoscaler object for the scaled object. As a result, the Horizontal Pod Autoscaler (HPA) Operator starts monitoring the scaling data associated with the trigger. The HPA requests scaling data from the cluster OpenShift API server endpoint.
The OpenShift API server endpoint is served by the custom metrics autoscaler metrics adapter. When the metrics adapter receives a request for custom metrics, it uses a GRPC connection to the controller to request it for the most recent trigger data received from the scaler.
The HPA makes scaling decisions based upon the data received from the metrics adapter and scales the workload up or down by increasing or decreasing the replicas.
As a it operates, a workload can affect the scaling metrics. For example, if a workload is scaled up to handle work in a Kafka queue, the queue size decreases after the workload processes all the work. As a result, the workload is scaled down.
If the metrics are in a range specified by the minReplicaCount value, the custom metrics autoscaler controller disables all scaling, and leaves the replica count at a fixed level. If the metrics exceed that range, the custom metrics autoscaler controller enables scaling and allows the HPA to scale the workload. While scaling is disabled, the HPA does not take any action.