Autoscaling with Cluster Agent Custom & External Metrics

Overview

Horizontal Pod Autoscaling, introduced in Kubernetes v1.2, allows autoscaling off of basic metrics like CPU, but it requires a resource called metrics-server to run alongside your application. As of Kubernetes v1.6, it is possible to autoscale off of custom metrics.

Custom metrics are user defined and are collected from within the cluster. As of Kubernetes v1.10, support for external metrics was introduced to autoscale off any metric from outside the cluster, such as those collected by Datadog.

You must first register the Cluster Agent as the External Metrics Provider. Next, adapt your HPAs to rely on the Cluster Agent’s provided metrics.

As of v1.0.0, the Custom Metrics Server in the Datadog Cluster Agent implements the External Metrics Provider interface for external metrics. This page explains how to set it up and how to autoscale your Kubernetes workload based on your Datadog metrics.

Setup

Requirements

  1. Kubernetes >v1.10: you must register the External Metrics Provider resource against the API server.
  2. Enable the Kubernetes aggregation layer.
  3. A valid Datadog API Key and Application Key.

Installation

To enable the external metrics server with your Cluster Agent managed by the Datadog Operator, first set up the Datadog Operator. Then, provide a valid Datadog API Key, Application Key, and set the features.externalMetricsServer.enabled to true in your DatadogAgent custom resource:

apiVersion: datadoghq.com/v2alpha1
kind: DatadogAgent
metadata:
  name: datadog
spec:
  global:
    credentials:
      apiKey: <DATADOG_API_KEY>
      appKey: <DATADOG_APP_KEY>

  features:
    externalMetricsServer:
      enabled: true

The Operator automatically updates the necessary RBAC configurations and sets the corresponding Service and APIService for Kubernetes to use.

The keys can alternatively be set by referencing the names of pre-created Secrets and the data keys storing your Datadog API and Application Keys.

apiVersion: datadoghq.com/v2alpha1
kind: DatadogAgent
metadata:
  name: datadog
spec:
  global:
    credentials:
      apiSecret:
        secretName: <SECRET_NAME>
        keyName: <KEY_FOR_DATADOG_API_KEY>
      appSecret:
        secretName: <SECRET_NAME>
        keyName: <KEY_FOR_DATADOG_APP_KEY>

  features:
    externalMetricsServer:
      enabled: true

To enable the external metrics server with your Cluster Agent in Helm, update your datadog-values.yaml file with the following configurations. Provide a valid Datadog API Key and Application Key, and set clusterAgent.metricsProvider.enabled to true. Then redeploy your Datadog Helm chart:

datadog:
  apiKey: <DATADOG_API_KEY>
  appKey: <DATADOG_APP_KEY>
  #(...)

clusterAgent:
  enabled: true
  # Enable the metricsProvider to be able to scale based on metrics in Datadog
  metricsProvider:
    # clusterAgent.metricsProvider.enabled
    # Set this to true to enable Metrics Provider
    enabled: true

This automatically updates the necessary RBAC configurations and sets up the corresponding Service and APIService for Kubernetes to use.

The keys can alternatively be set by referencing the names of pre-created Secrets containing the data keys api-key and app-key with the configurations datadog.apiKeyExistingSecret and datadog.appKeyExistingSecret.

Custom metrics server

To enable the Custom Metrics Server, first follow the instructions to set up the Datadog Cluster Agent within your cluster. Once you have verified a successful base deployment, edit your Deployment manifest for the Datadog Cluster Agent with the following steps:

  1. Set DD_EXTERNAL_METRICS_PROVIDER_ENABLED environment variable to true.
  2. Ensure you have both your environment variables DD_APP_KEY and DD_API_KEY set.
  3. Ensure you have your DD_SITE environment variable set to your Datadog site: . It defaults to the US site datadoghq.com.

Register the external metrics provider service

Once the Datadog Cluster Agent is up and running, apply some additional RBAC policies and set up the Service to route the corresponding requests.

  1. Create a Service named datadog-custom-metrics-server, exposing the port 8443 with the following custom-metric-server.yaml manifest:

    kind: Service
    apiVersion: v1
    metadata:
      name: datadog-custom-metrics-server
    spec:
      selector:
        app: datadog-cluster-agent
      ports:
      - protocol: TCP
        port: 8443
        targetPort: 8443
    

    Note: The Cluster Agent by default is expecting these requests over port 8443. However, if your Cluster Agent Deployment has set the environment variable DD_EXTERNAL_METRICS_PROVIDER_PORT to some other port value, change the targetPort of this Service accordingly.

    Apply this Service by running kubectl apply -f custom-metric-server.yaml

  2. Download the rbac-hpa.yaml RBAC rules file.

  3. Register the Cluster Agent as an external metrics provider by applying this file:

    kubectl apply -f rbac-hpa.yaml
    

Once enabled, the Cluster Agent is ready to fetch metrics for the HPA. There are two options:

Datadog recommends using the DatadogMetric option. While this does require an additional step of deploying the DatadogMetric CustomResourceDefinition (CRD), this provides a lot more flexibility in the queries performed. If you do not use DatadogMetric queries, your HPAs use the native Kubernetes external metrics format, which the Cluster Agent translates into a Datadog metric query.

If you are dual-shipping your metrics to multiple Datadog organizations, you can configure the Cluster Agent to fetch from these multiple endpoints for high availability. For more information, see the Dual Shipping documentation.

Autoscaling with DatadogMetric queries

You can autoscale on a Datadog query by using the DatadogMetric Custom Resource Definition (CRD) and Datadog Cluster Agent versions 1.7.0 or above. This is a more flexible approach and allows you to scale with the exact Datadog query you would use in-app.

Requirements

For autoscaling to work correctly, custom queries must follow these rules:

  • The query must be syntactically correct, otherwise it prevents the refresh of ALL metrics used for autoscaling (effectively stopping autoscaling).
  • The query result must output only one series (otherwise, the results are considered invalid).
  • The query should yield at least two non-null timestamped points (it’s possible to use a query that returns a single point, though in this case, autoscaling may use incomplete points).

Note: While the query is arbitrary, the start and end times are still set at Now() - 5 minutes and Now() by default.

Setup DatadogMetric CRD

The Custom Resource Definition (CRD) for the DatadogMetric object can be added to your Kubernetes cluster by using Helm, the Datadog Operator, or Daemonset:

To activate the usage of the DatadogMetric CRD update your DatadogAgent custom resource and set features.externalMetricsServer.useDatadogMetrics to true.

kind: DatadogAgent
apiVersion: datadoghq.com/v2alpha1
metadata:
  name: datadog
spec:
  global:
    credentials:
      apiKey: <DATADOG_API_KEY>
      appKey: <DATADOG_APP_KEY>
  features:
    externalMetricsServer:
      enabled: true
      useDatadogMetrics: true

The Operator automatically updates the necessary RBAC configurations and directs the Cluster Agent to manage these HPA queries through these DatadogMetric resources.

To activate usage of the DatadogMetric CRD update your values.yaml Helm configuration to set clusterAgent.metricsProvider.useDatadogMetrics to true. Then redeploy your Datadog Helm chart:

clusterAgent:
  enabled: true
  metricsProvider:
    enabled: true
    # clusterAgent.metricsProvider.useDatadogMetrics
    # Enable usage of DatadogMetric CRD to autoscale on arbitrary Datadog queries
    useDatadogMetrics: true

Note: This attempts to install the DatadogMetric CRD automatically. If that CRD already exists prior to the initial Helm installation, it may conflict.

This automatically updates the necessary RBAC files and directs the Cluster Agent to manage these HPA queries through these DatadogMetric resources.

To activate usage of the DatadogMetric CRD, follow these extra steps:

  1. Install the DatadogMetric CRD in your cluster.

    kubectl apply -f "https://raw.githubusercontent.com/DataDog/helm-charts/master/crds/datadoghq.com_datadogmetrics.yaml"
    
  2. Update Datadog Cluster Agent RBAC manifest, it has been updated to allow usage of DatadogMetric CRD.

    kubectl apply -f "https://raw.githubusercontent.com/DataDog/datadog-agent/master/Dockerfiles/manifests/cluster-agent-datadogmetrics/cluster-agent-rbac.yaml"
    
  3. Set the DD_EXTERNAL_METRICS_PROVIDER_USE_DATADOGMETRIC_CRD to true in the deployment of the Datadog Cluster Agent.

Create the DatadogMetric object

Once the DatadogMetric custom resource has been added to your cluster, you can create DatadogMetric objects for your HPAs to reference. While any HPA can reference any DatadogMetric, Datadog recommends creating them in the same namespace as your HPA.

Note: Multiple HPAs can use the same DatadogMetric.

You can create a DatadogMetric with the following manifest:

apiVersion: datadoghq.com/v1alpha1
kind: DatadogMetric
metadata:
  name: <DATADOG_METRIC_NAME>
spec:
  query: <CUSTOM_QUERY>

Example DatadogMetric object

For example, a DatadogMetric object to autoscale an NGINX deployment based on the nginx.net.request_per_s Datadog metric:

apiVersion: datadoghq.com/v1alpha1
kind: DatadogMetric
metadata:
  name: nginx-requests
spec:
  query: max:nginx.net.request_per_s{kube_container_name:nginx}.rollup(60)

Use DatadogMetric in HPA

Once your Cluster Agent has been set up and DatadogMetric created, update your HPA to reference this DatadogMetric relative to its namespace and name. The general format is to specify the metric for the HPA as a type: External and specify the metric name like datadogmetric@<NAMESPACE>:<DATADOG_METRIC_NAME>.

Example HPAs with DatadogMetric

An HPA using the DatadogMetric named nginx-requests, assuming both objects are in namespace nginx-demo.

Using apiVersion: autoscaling/v2:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: nginxext
spec:
  minReplicas: 1
  maxReplicas: 3
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: nginx
  metrics:
  - type: External
    external:
      metric:
        name: datadogmetric@nginx-demo:nginx-requests
      target:
        type: Value
        value: 9

Using apiVersion: autoscaling/v2beta1:

apiVersion: autoscaling/v2beta1
kind: HorizontalPodAutoscaler
metadata:
  name: nginxext
spec:
  minReplicas: 1
  maxReplicas: 3
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: nginx
  metrics:
  - type: External
    external:
      metricName: datadogmetric@nginx-demo:nginx-requests
      targetValue: 9

In these manifests:

  • The HPA is configured to autoscale the deployment called nginx.
  • The maximum number of replicas created is 3, and the minimum is 1.
  • The HPA relies on the DatadogMetric nginx-requests in the nginx-demo namespace.

Once the DatadogMetric is linked to an HPA, the Datadog Cluster Agent marks it as active. The Cluster Agent then makes requests to Datadog for the query, stores the results in the DatadogMetric object, and provides the values to the HPA.

Autoscaling without DatadogMetric queries

If you do not want to autoscale with the DatadogMetric you can still create your HPAs with the native Kubernetes format. The Cluster Agent converts the HPA format into a Datadog metric query.

Once you have the Datadog Cluster Agent running and the service registered, create an HPA manifest and specify type: External for your metrics. This notifies the HPA to pull the metrics from the Datadog Cluster Agent’s service:

spec:
  metrics:
    - type: External
      external:
        metricName: "<METRIC_NAME>"
        metricSelector:
          matchLabels:
            <TAG_KEY>: <TAG_VALUE>

Example HPAs without DatadogMetric

An HPA manifest to autoscale off an NGINX deployment based off of the nginx.net.request_per_s Datadog metric using apiVersion: autoscaling/v2:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: nginxext
spec:
  minReplicas: 1
  maxReplicas: 3
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: nginx
  metrics:
  - type: External
    external:
      metric:
        name: nginx.net.request_per_s
      target:
        type: Value
        value: 9

The following is the same HPA manifest as above using apiVersion: autoscaling/v2beta1:

apiVersion: autoscaling/v2beta1
kind: HorizontalPodAutoscaler
metadata:
  name: nginxext
spec:
  minReplicas: 1
  maxReplicas: 3
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: nginx
  metrics:
  - type: External
    external:
      metricName: nginx.net.request_per_s
      metricSelector:
        matchLabels:
            kube_container_name: nginx
      targetValue: 9

In these manifests:

  • The HPA is configured to autoscale the deployment called nginx.
  • The maximum number of replicas created is 3, and the minimum is 1.
  • The metric used is nginx.net.request_per_s, and the scope is kube_container_name: nginx. This format corresponds to Datadog’s metrics format.

Every 30 seconds, Kubernetes queries the Datadog Cluster Agent to get the value of this metric and autoscales proportionally if necessary. For advanced use cases, it is possible to have several metrics in the same HPA. As noted in Kubernetes horizontal pod autoscaling, the largest of the proposed values is the one chosen.

Migration

Existing HPAs are automatically migrated using external metrics.

When you set DD_EXTERNAL_METRICS_PROVIDER_USE_DATADOGMETRIC_CRD to true but you still have HPAs that do not reference a DatadogMetric, normal syntax (without referencing a DatadogMetric through datadogmetric@...) is still supported.

The Datadog Cluster Agent automatically creates DatadogMetric resources in its own namespace (their name starts with dcaautogen-) to accommodate this, allowing a smooth transition to DatadogMetric.

If you choose to migrate an HPA later on to reference a DatadogMetric, the automatically generated resource is cleaned up by the Datadog Cluster Agent after a few hours.

Cluster Agent querying

The Cluster Agent performs the queries for the DatadogMetric objects every 30 seconds. The Cluster Agent also batches the metric queries performed into groups of 35. Therefore, 35 DatadogMetric queries are included in a single request to the Datadog metrics API.

By batching these queries, the Cluster Agent can perform them more efficiently and avoid being rate-limited.

This means that the Cluster Agent performs roughly 120 API requests per hour per 35 DatadogMetric objects. As you add more DatadogMetric objects or add the autoscaling functionality to additional Kubernetes clusters, this increases the number of calls to fetch metrics within the same org.

The Cluster Agent also queries for the past five minutes of data by default for each of these metric queries. This is to ensure the Cluster Agent is scaling off recent data. However, if your metric queries are relying on data from one of the cloud integrations (AWS, Azure, GCP, etc.), these are fetched at a slight delay and are not covered within the five minute interval. In these cases, provide the environment variables to the Cluster Agent to increase the date range and data age allowed for the metric queries.

- name: DD_EXTERNAL_METRICS_PROVIDER_BUCKET_SIZE
  value: "900"
- name: DD_EXTERNAL_METRICS_PROVIDER_MAX_AGE
  value: "900"

Troubleshooting

DatadogMetric Status

The Datadog Cluster Agent takes care of updating the status subresource of all DatadogMetric resources to reflect results from queries to Datadog. This is the main source of information to understand what happens if something is failing. You can run the following to get this information outputted for you:

kubectl describe datadogmetric <RESOURCE NAME>

Example

The status part of a DatadogMetric:

status:
  conditions:
  - lastTransitionTime: "2020-06-22T14:38:21Z"
    lastUpdateTime: "2020-06-25T09:21:00Z"
    status: "True"
    type: Active
  - lastTransitionTime: "2020-06-25T09:00:00Z"
    lastUpdateTime: "2020-06-25T09:21:00Z"
    status: "True"
    type: Valid
  - lastTransitionTime: "2020-06-22T14:38:21Z"
    lastUpdateTime: "2020-06-25T09:21:00Z"
    status: "True"
    type: Updated
  - lastTransitionTime: "2020-06-25T09:00:00Z"
    lastUpdateTime: "2020-06-25T09:21:00Z"
    status: "False"
    type: Error
  currentValue: "1977.2"

The four conditions give you insights on the current state of your DatadogMetric:

  • Active: Datadog considers a DatadogMetric active if at least one HPA is referencing it. Inactive DatadogMetrics are not updated to minimize API usage.
  • Valid: Datadog considers a DatadogMetric valid when the answer for the associated query is valid. An invalid status probably means that your custom query is not semantically correct. See the Error field for details.
  • Updated: This condition is always updated when the Datadog Cluster Agent touches a DatadogMetric.
  • Error: If processing this DatadogMetric triggers an error, this condition is true and contains error details.

The currentValue is the value gathered from Datadog and returned to the HPAs.

Value vs AverageValue for the target metric

The HPAs in these example use the target type of Value instead of AverageValue. Both options are supported. Adjust your Datadog metric queries accordingly.

When using Value, the metric value returned by the Datadog metric query is provided exactly as-is to the HPA for its scaling decision. When using AverageValue, the metric value returned is divided by the current number of pods. Set your <Metric Value> accordingly to how you want your HPA to scale based on your query and returned value.

Using apiVersion: autoscaling/v2, the target metric configuration for Value looks like:

  metrics:
  - type: External
    external:
      metric:
        name: datadogmetric@<NAMESPACE>:<DATADOG_METRIC_NAME>
      target:
        type: Value
        value: <METRIC_VALUE>

Whereas AverageValue looks like:

  metrics:
  - type: External
    external:
      metric:
        name: datadogmetric@<NAMESPACE>:<DATADOG_METRIC_NAME>
      target:
        type: AverageValue
        averageValue: <METRIC_VALUE>

For apiVersion: autoscaling/v2beta1 the respective options are targetValue and targetAverageValue.

Further Reading

PREVIEWING: may/unit-testing