Autoscaling in a Nutshell

Kubernetes workload resources (such as the Cloudentity platform deployment) can be automatically scaled to match demand based on custom metrics, multiple metrics, or metric APIs. For example, increased load may result in additional pods being deployed. On the other hand, if the load decreases, the workload resources is scaled back down.

Cloudentity Helm Chart provides support for the HorizontalPodAutoscaler resource. Metrics Server must be used to expose Cloudentity resources usage metrics used by autoscaling.

More on autoscaling details can be found in the Horizontal Pod Autoscaler K8s documentation.

Prerequisites

• Kubernetes cluster v1.16+
• Kubernetes Metrics Server
• Helm v3.0+
• Resource requests specified in the Helm chart

Configure Autoscaling

Autoscaling can be enabled for base Cloudentity pods as well as worker pods. The configuration parameters are identical, although worker pods configuration is located under the workers key in values.yaml.

For autoscaling to work properly, resource requests must be set for the Cloudentity pods.

resources:
  requests:
    cpu: 500m
    memory: 1.2Gi

To enable autoscaling integration, set the autoscaling.enabled parameter to true.

When autoscaling is enabled, the replicaCount parameter is ignored.

• autoscaling.minReplicas parameter is used to set the minimum number of replicas.
• autoscaling.maxReplicas parameter is used to set the maximum number of replicas.
• autoscaling.targetCPUUtilizationPercentage parameter can be used to enable CPU autoscaling at a given percentage.
• autoscaling.targetMemoryUtilizationPercentage parameter can be used to enable memory autoscaling at a given percentage.
• autoscaling.behavior parameter can be used to configure detailed scaling behaviors

autoscaling:
  ## If true, autoscaling is enabled
  ##
  enabled: true

  ## Set a minimum number of 3 replicas
  ##
  minReplicas: 3

  ## Set a maximum number of 9 replicas
  ##
  maxReplicas: 9

  ## Enable CPU autoscaling at 70% of request utilization
  ##
  targetCPUUtilizationPercentage: 70

  ## Enable memory autoscaling at 50% of request utilization
  ##
  targetMemoryUtilizationPercentage: 50

  ## Consider utilization values from last 5 minutes during scaling
  ## Scale Down one pod at a time every 180 seconds
  ##
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 300
      policies:
        - type: Pods
          value: 1
          periodSeconds: 180

Enhancing Cloudentity Autoscaling

This section delves into the more intricate aspects of ACP autoscaling. The concepts outlined here may be beneficial and can be adapted according to your requirements.

Kubernetes Event-Driven Autoscaling for Cloudentity workers

Kubernetes Event-Driven Autoscaling (KEDA) is an open-source project developed in collaboration between Microsoft and Red Hat. KEDA is an event-driven autoscaler that allows applications to scale based on events from various event sources, such as Prometheus, Kafka, RabbitMQ, and many more. KEDA provides a Kubernetes-native way to scale applications in response to events, rather than traditional metrics like CPU or memory usage.

In scenarios where an exceptionally high number of events are anticipated in Cloudentity, there’s a risk that the handlers may not keep pace and the HPA may not adequately scale up the Cloudentity worker pods. This situation could lead to events being processed with a significant delay, potentially a few minutes. Should you encounter such a scenario, we suggest leveraging KEDA. It can drive the scaling up of Cloudentity workers HPA based on the lag of the stream handler rather than relying on CPU usage.

Configuration

1. Install monitoring stack for your Cloudentity installation (docs).

2. Install KEDA in your Kubernetes cluster from the offical Helm chart (docs).

   Below is some basic configuration for the chart we recommend.

   operator:
     replicaCount: 2
     affinity:
       podAntiAffinity:
         requiredDuringSchedulingIgnoredDuringExecution:
         - labelSelector:
             matchExpressions:
             - key: app
               operator: In
               values:
               - keda-operator
           topologyKey: "kubernetes.io/hostname"
   metricsServer:
     replicaCount: 2
     affinity:
       podAntiAffinity:
         requiredDuringSchedulingIgnoredDuringExecution:
         - labelSelector:
             matchExpressions:
             - key: app
               operator: In
               values:
               - keda-operator-metrics-apiserver
           topologyKey: "kubernetes.io/hostname"
   tolerations:
     - key: CriticalAddonsOnly
       value: "true"
       effect: NoSchedule
   topologySpreadConstraints:
     operator:
       - labelSelector:
           matchLabels:
             app: keda-operator
         maxSkew: 1
         topologyKey: topology.kubernetes.io/zone
         whenUnsatisfiable: ScheduleAnyway
     metricsServer:
       - labelSelector:
           matchLabels:
             app: keda-operator-metrics-apiserver
         maxSkew: 1
         topologyKey: topology.kubernetes.io/zone
         whenUnsatisfiable: ScheduleAnyway
   

   For more parameters refer to the official values.yaml description.

3. Create KEDA ScaledObject based on the Cloudentity acp_redis_lag_seconds_bucket metric.

   Below is our recommendation of how it can be configured.

   apiVersion: keda.sh/v1alpha1
   kind: ScaledObject
   metadata:
     name: acp-workers
   spec:
     scaleTargetRef:
       name: acp-workers
     minReplicaCount: 1
     maxReplicaCount: 1
     triggers:
     - type: prometheus
       metadata:
         name: prom-trigger
         serverAddress: <prometheus endpoint cluster url>
         threshold: '10'
         query: 'round(max(avg(histogram_quantile(0.9, rate(acp_redis_lag_seconds_bucket{job="acp-workers", namespace="acp"}[1m])) > 0) by (group, stream)))'
     advanced:
       horizontalPodAutoscalerConfig:
         behavior:
           scaleDown:
             stabilizationWindowSeconds: 300
             policies:
             - type: Pods
               value: 2
               periodSeconds: 100
           scaleUp:
             stabilizationWindowSeconds: 60
             policies:
             - type: Pods
               value: 4
               periodSeconds: 30
   

   The PromQL query defined in spec.triggers.0.metadata.query field calculates average 90th percentile of Redis lag across all ACP workers pods over the last 1 min period. If the calculated value is higher than spec.triggers.0.metadata.threshold value, KEDA will start the scaling-up procedure.

4. Disable the default Horizontal Pod Autoscaling on the Cloudentity workers by switching the workers.autoscaling.enabled value to false in the main Cloudentity configuration yaml:

   workers:
     autoscaling:
       enabled: false
   

   Cloudentity is not needed as KEDA will take care of the autoscaling process of Cloudentity worker pods since now.

Conclusion

• Autoscaling works based on the average usage across all Cloudentity pods for given metric.
• Average value of 100% for the metric is defined in resources.requests.
• CPU and Memory autoscaling can work at the same time. Scaling works on metric which reports the higher desired count.
• The default upscale delay equals 0s (controlled by cluster operator in kube-controller-manager)
• The default downscale delay equals 5m (controlled by cluster operator in kube-controller-manager)
• Metrics scrape interval can be configured in metric-server via the metric-resolution (default 60s)
• The default scale up bahavior is to add 100% of currently running replicas or 4 pods (whichever is higher) every 15 seconds based on last metric
• The default scale down bahavior is to remove up to 100% of currently running replicas every 15 seconds based on 5 minutes of metrics