Scale and autoscale runtime services

You are currently viewing version 1.11 of the Apigee hybrid documentation. For more information, see Supported versions.

Version 1.11 

   Supported versions:

• v1.14 (latest)                        
• v1.13                         
• v1.12                         
• v1.11                         
• List of supported versions

   Unsupported versions:

• v1.10                                      
• v1.9                                         
• v1.8                                         
• v1.7                                         
• v1.6                                         
• v1.5                                         
• v1.4                                         
• v1.3                                         
• v1.2                                         
• v1.1                                         

You can scale most services running in Kubernetes from the command line or in a configuration override. You can set scaling parameters for Apigee hybrid runtime services in the overrides.yaml file.

Service	Implemented As	Scaling
Cassandra	ApigeeDatastore (CRD)	See Scaling Cassandra.
Ingress/LoadBalancer	Deployment	Anthos Service Mesh uses Horizontal Pod Autoscaling (HPAs).
Logger	DaemonSet	DaemonSets manage replicas of a pod on all nodes, so they scale when you scale the pods themselves.
MART Apigee Connect Watcher	ApigeeOrganization (CRD)	To scale via configuration, increase the value of the Deployment's replicaCountMin configuration property for the mart, watcher, and/or connectAgent stanzas. For example: mart: replicaCountMax: 2 replicaCountMin: 1 watcher: replicaCountMax: 2 replicaCountMin: 1 connectAgent: replicaCountMax: 2 replicaCountMin: 1 These Deployments use a Horizontal Pod Autoscaler for autoscaling. Set the Deployment object's targetCPUUtilizationPercentage property to the threshold for scaling up; when this value is exceeded, Kubernetes adds pods up to the value of replicaCountMax. For more information on setting configuration properties, see Manage runtime plane components.
Runtime Synchronizer UDCA	ApigeeEnvironment (CRD)	To scale via configuration, increase the value of the replicaCountMin property for the udca, synchronizer, and/or runtime stanzas in the overrides file. For example: synchronizer: replicaCountMax: 10 replicaCountMin: 1 runtime: replicaCountMax: 10 replicaCountMin: 1 udca: replicaCountMax: 10 replicaCountMin: 1 Note: These changes apply to ALL environments in the overrides file. If you wish to customize scaling for each environment see Advanced configurations below. These deployments use a Horizontal Pod Autoscaler for autoscaling. Set the Deployment object's targetCPUUtilizationPercentage property to the threshold for scaling up; when this value is exceeded, Kubernetes adds pods up to the value of replicaCountMax. For more information on setting configuration properties, see Manage runtime plane components.

Advanced configurations

In some scenarios, you may need to use advanced scaling options. Example scenarios include:

• Setting different scaling options for each environment. For example, where env1 has a minReplica of 5 and env2 has a minReplica of 2.
• Setting different scaling options for each component within an environment. For example, where the udca component has a maxReplica of 5 and the synchronizer component has a maxReplica of 2.

The following example shows how to use the kubernetes patch command to change the maxReplicas property for the runtime component:

1. Create environment variables to use with the command:

   export ENV=my-environment-name
   export NAMESPACE=apigee  #the namespace where apigee is deployed
   export COMPONENT=runtime #can be udca or synchronizer
   export MAX_REPLICAS=2
   export MIN_REPLICAS=1

2. Apply the patch. Note that this example assumes that kubectl is in your PATH:

   kubectl patch apigeeenvironment -n $NAMESPACE \
     $(kubectl get apigeeenvironments -n $NAMESPACE -o jsonpath='{.items[?(@.spec.name == "'$ENV'" )]..metadata.name}') \
     --patch "$(echo -e "spec:\n  components:\n    $COMPONENT:\n      autoScaler:\n        maxReplicas: $MAX_REPLICAS\n        minReplicas: $MIN_REPLICAS")" \
     --type merge

3. Verify the change:

   kubectl get hpa -n $NAMESPACE

Environment-based scaling

By default, scaling is described at the organization level. You can override the default settings by specifying environment-specific scaling in the overrides.yaml file as shown in the following example:

envs:
  # Apigee environment name
  - name: test
    components:
    # Environment-specific scaling override
    # Otherwise, uses scaling defined at the respective root component
     runtime:
      replicaCountMin: 2
      replicaCountMax: 20
  

Metrics-based scaling

With metrics-based scaling, the runtime can use CPU and application metrics to scale the apigee-runtime pods. The Kubernetes Horizontal Pod Autoscaler (HPA) API, uses the hpaBehavior field to configure the scale-up and scale-down behaviors of the target service. Metrics-based scaling is not available for any other components in a hybrid deployment.

Note: An internal connection between Prometheus and the Prometheus-Adapter on port 6443 must be open in order to receive metrics data and enable scaling. For more information on required ports, see Internal connections.

Scaling can be adjusted based on the following metrics:

Metric	Measure	Considerations
serverNioTaskWaitTime	Average wait time (in picoseconds) of processing queue in runtime instances for proxy requests at the http layer.	This metric measures the impact of the number and payload size of proxy requests and responses.
serverMainTaskWaitTime	Average wait time (in picoseconds) of processing queue in runtime instances for proxy requests to process policies.	This metric measures the impact of complexity in the policies attached to the proxy request flow.

The following example from the runtime stanza in the overrides.yaml illustrates the standard parameters (and permitted ranges) for scaling apigee-runtime pods in a hybrid implementation:

hpaMetrics:
      serverMainTaskWaitTime: 400M (300M to 450M)
      serverNioTaskWaitTime: 400M (300M to 450M)
      targetCPUUtilizationPercentage: 75
    hpaBehavior:
      scaleDown:
        percent:
          periodSeconds: 60 (30 - 180)
          value: 20 (5 - 50)
        pods:
          periodSeconds: 60 (30 - 180)
          value: 2 (1 - 15)
        selectPolicy: Min
        stabilizationWindowSeconds: 120 (60 - 300)
      scaleUp:
        percent:
          periodSeconds: 60 (30 - 120)
          value: 20 (5 - 100)
        pods:
          periodSeconds: 60 (30 - 120)
          value: 4 (2 - 15)
        selectPolicy: Max
        stabilizationWindowSeconds: 30 (30 - 120)
  

Configure more aggressive scaling

Increasing the percent and pods values of the scale-up policy will result in a more aggressive scale-up policy. Similarly, increasing the percent and pods values in scaleDown will result in an aggressive scale-down policy. For example:

hpaMetrics:
    serverMainTaskWaitTime: 400M
    serverNioTaskWaitTime: 400M
    targetCPUUtilizationPercentage: 75
  hpaBehavior:
    scaleDown:
      percent:
        periodSeconds: 60
        value: 20
      pods:
        periodSeconds: 60
        value: 4
      selectPolicy: Min
      stabilizationWindowSeconds: 120
    scaleUp:
      percent:
        periodSeconds: 60
        value: 30
      pods:
        periodSeconds: 60
        value: 5
      selectPolicy: Max
      stabilizationWindowSeconds: 30

In the above example, the scaleDown.pods.value is increased to 5, the scaleUp.percent.value is increased to 30, and the scaleUp.pods.value is increased to 5.

Note: The value of periodSeconds should not go below 30.

Configure less aggressive scaling

The hpaBehavior configuration values can also be decreased to implement less aggressive scale-up and scale-down policies. For example:

hpaMetrics:
    serverMainTaskWaitTime: 400M
    serverNioTaskWaitTime: 400M
    targetCPUUtilizationPercentage: 75
  hpaBehavior:
    scaleDown:
      percent:
        periodSeconds: 60
        value: 10
      pods:
        periodSeconds: 60
        value: 1
      selectPolicy: Min
      stabilizationWindowSeconds: 180
    scaleUp:
      percent:
        periodSeconds: 60
        value: 20
      pods:
        periodSeconds: 60
        value: 4
      selectPolicy: Max
      stabilizationWindowSeconds: 30

In the above example, the scaleDown.percent.value is decreased to 10, the scaleDown.pods.value is decreased to 1, and the scaleUp.stablizationWindowSeconds is increased to 180.

For more information about metrics-based scaling using the hpaBehavior field, see Scaling policies.

Disable metrics-based scaling

While metrics-based scaling is enabled by default and cannot be completely disabled, you can configure the metrics thresholds at a level that metrics-based scaling will not be triggered. The resulting scaling behavior will be the same as CPU-based scaling. For example, you can use the following configuration to prevent triggering metrics-based scaling:

hpaMetrics:
      serverMainTaskWaitTime: 4000M
      serverNioTaskWaitTime: 4000M
      targetCPUUtilizationPercentage: 75
    hpaBehavior:
      scaleDown:
        percent:
          periodSeconds: 60
          value: 10
        pods:
          periodSeconds: 60
          value: 1
        selectPolicy: Min
        stabilizationWindowSeconds: 180
      scaleUp:
        percent:
          periodSeconds: 60
          value: 20
        pods:
          periodSeconds: 60
          value: 4
        selectPolicy: Max
        stabilizationWindowSeconds: 30

Troubleshooting

This section describes troubleshooting methods for common errors you may encounter while configuring scaling and auto-scaling.

HPA shows unknown for metrics values

If metrics-based scaling does not work and the HPA shows unknown for metrics values, use the following command to check the HPA output:

kubectl describe hpa HPA_NAME

When running the command, replace HPA_NAME with the name of the HPA you wish to view.

The output will show the CPU target and utilization of the service, indicating that CPU scaling will work in the absence of metrics-based scaling. For HPA behavior using multiple parameters, see Scaling on multiple metrics.