Deployments autoscaling on ECK

Note

Elasticsearch autoscaling requires a valid Enterprise license or Enterprise trial license. Check the license documentation for more details about managing licenses.

ECK can leverage the autoscaling API introduced in Elasticsearch 7.11 to adjust automatically the number of Pods and the allocated resources in a tier. Currently, autoscaling is supported for Elasticsearch data tiers and machine learning nodes.

Enable autoscaling

To enable autoscaling on an Elasticsearch cluster, you need to define one or more autoscaling policies. Each autoscaling policy applies to one or more NodeSets which share the same set of roles specified in the node.roles setting in the Elasticsearch configuration.

Define autoscaling policies

Autoscaling policies can be defined in an ElasticsearchAutoscaler resource. Each autoscaling policy must have the following fields:

• name is a unique name used to identify the autoscaling policy.
• roles contains a set of node roles, unique across all the autoscaling policies, used to identify the NodeSets to which this policy applies. At least one NodeSet with the exact same set of roles must exist in the Elasticsearch resource specification.
• resources helps define the minimum and maximum compute resources usage:
  • nodeCount defines the minimum and maximum nodes allowed in the tier.
  • cpu and memory enforce minimum and maximum compute resources usage for the Elasticsearch container.
  • storage enforces minimum and maximum storage request per PersistentVolumeClaim.

If there is no recommendation from the Autoscaling API for a given resource, and if this resource is not set in the policy, then the resource is not managed by the operator and existing requirements in the NodeSets remain unchanged.

			apiVersion: autoscaling.k8s.elastic.co/v1alpha1
kind: ElasticsearchAutoscaler
metadata:
  name: autoscaling-sample
spec:
  ## The name of the Elasticsearch cluster to be scaled automatically.
  elasticsearchRef:
    name: elasticsearch-sample
  ## The autoscaling policies.
  policies:
    - name: data-ingest
      roles: ["data", "ingest" , "transform"]
      resources:
        nodeCount:
          min: 3
          max: 8
        cpu:
          min: 2
          max: 8
        memory:
          min: 2Gi
          max: 16Gi
        storage:
          min: 64Gi
          max: 512Gi
    - name: ml
      roles:
        - ml
      resources:
        nodeCount:
          min: 1
          max: 9
        cpu:
          min: 1
          max: 4
        memory:
          min: 2Gi
          max: 8Gi
        storage:
          min: 1Gi
          max: 1Gi

		

Warning

A node role should not be referenced in more than one autoscaling policy.

In the case of storage the following restrictions apply:

• Scaling the storage size automatically requires the ExpandInUsePersistentVolumes feature to be enabled. It also requires a storage class that supports volume expansion.
• Only one persistent volume claim per Elasticsearch node is supported when autoscaling is enabled.
• Volume size cannot be scaled down.
• Scaling up (vertically) is only supported if the available capacity in a PersistentVolume matches the capacity claimed in the PersistentVolumeClaim. Refer to the next section for more information.

Scale Up and Scale Out

In order to adapt the resources to the workload, the operator first attempts to scale up the resources (cpu, memory, and storage) allocated to each node in the NodeSets. The operator always ensures that the requested resources are within the limits specified in the autoscaling policy. If each individual node has reached the limits specified in the autoscaling policy, but more resources are required to handle the load, then the operator adds some nodes to the NodeSets. Nodes are added up to the max value specified in the nodeCount of the policy.

Warning

Scaling up (vertically) is only supported if the actual storage capacity of the persistent volumes matches the capacity claimed. If the physical capacity of a PersistentVolume may be greater than the capacity claimed in the PersistentVolumeClaim, it is advised to set the same value for the min and the max setting of each resource. It is however still possible to let the operator scale out the NodeSets automatically, as in the following example:

			apiVersion: autoscaling.k8s.elastic.co/v1alpha1
kind: ElasticsearchAutoscaler
metadata:
  name: autoscaling-sample
spec:
  elasticsearchRef:
    name: elasticsearch-sample
  policies:
    - name: data-ingest
      roles: ["data", "ingest" , "transform"]
      resources:
        nodeCount:
          min: 3
          max: 9
        cpu:
          min: 4
          max: 4
        memory:
          min: 16Gi
          max: 16Gi
        storage:
          min: 512Gi
          max: 512Gi

		

Set the limits

The value set for memory and CPU limits are computed by applying a ratio to the calculated resource request. The default ratio between the request and the limit for both CPU and memory is 1. This means that request and limit have the same value. You can change the default ratio between the request and the limit for both the CPU and memory ranges by using the requestsToLimitsRatio field.

For example, you can set a CPU limit to twice the value of the request, as follows:

			apiVersion: autoscaling.k8s.elastic.co/v1alpha1
kind: ElasticsearchAutoscaler
metadata:
  name: autoscaling-sample
spec:
  elasticsearchRef:
    name: elasticsearch-sample
  policies:
    - name: data-ingest
      roles: ["data", "ingest" , "transform"]
      resources:
        nodeCount:
          min: 2
          max: 5
        cpu:
          min: 1
          max: 2
          requestsToLimitsRatio: 2
        memory:
          min: 2Gi
          max: 6Gi
        storage:
          min: 512Gi
          max: 512Gi

		

You can find a complete example in the ECK GitHub repository which will also show you how to fine-tune the autoscaling deciders.

Change the polling interval

The Elasticsearch autoscaling capacity endpoint is polled every minute by the operator. This interval duration can be controlled using the pollingPeriod field in the autoscaling specification:

			apiVersion: autoscaling.k8s.elastic.co/v1alpha1
kind: ElasticsearchAutoscaler
metadata:
  name: autoscaling-sample
spec:
  pollingPeriod: "42s"
  elasticsearchRef:
    name: elasticsearch-sample
  policies:
    - name: data-ingest
      roles: ["data", "ingest" , "transform"]
      resources:
        nodeCount:
          min: 2
          max: 5
        cpu:
          min: 1
          max: 2
        memory:
          min: 2Gi
          max: 6Gi
        storage:
          min: 512Gi
          max: 512Gi

		

Monitoring

Autoscaling status

In addition to the logs generated by the operator, an autoscaling status is maintained in the ElasticsearchAutoscaler resource. This status holds several Conditions to summarize the health and the status of the autoscaling mechanism. For example, dedicated Conditions may report if the controller cannot connect to the Elasticsearch cluster, or if a resource limit has been reached:

			kubectl get elasticsearchautoscaler autoscaling-sample \
    -o jsonpath='{ .status.conditions }' | jq

		

			[
 {
   "lastTransitionTime": "2022-09-09T08:07:10Z",
   "message": "Limit reached for policies data-ingest",
   "status": "True",
   "type": "Limited"
 },
  {
   "lastTransitionTime": "2022-09-09T07:55:08Z",
   "status": "True",
   "type": "Active"
 },
 {
   "lastTransitionTime": "2022-09-09T08:07:10Z",
   "status": "True",
   "type": "Healthy"
 },
 {
   "lastTransitionTime": "2022-09-09T07:56:22Z",
   "message": "Elasticsearch is available",
   "status": "True",
   "type": "Online"
 }
]

		

Expected resources

The autoscaler status also contains a policies section which describes the expected resources for each NodeSet managed by an autoscaling policy.

			kubectl get elasticsearchautoscaler.autoscaling.k8s.elastic.co/autoscaling-sample \
    -o jsonpath='{ .status.policies }' | jq

		

			[
  {
    "lastModificationTime": "2022-10-05T05:47:13Z",
    "name": "data-ingest",
    "nodeSets": [
      {
        "name": "nodeset-1",
        "nodeCount": 2
      }
    ],
    "resources": {
      "limits": {
        "cpu": "1",
        "memory": "2Gi"
      },
      "requests": {
        "cpu": "500m",
        "memory": "2Gi",
        "storage": "1Gi"
      }
    }
  }
]

		

Events

Important events are also reported through Kubernetes events, for example when the maximum autoscaling size limit is reached:

			> kubectl get events

40m  Warning  HorizontalScalingLimitReached  elasticsearch/sample   Can't provide total required storage 32588740338, max number of nodes is 5, requires 6 nodes

		

Disable autoscaling

You can disable autoscaling at any time by deleting the ElasticsearchAutoscaler resource. For machine learning the following settings are not automatically reset:

• xpack.ml.max_ml_node_size
• xpack.ml.max_lazy_ml_nodes
• xpack.ml.use_auto_machine_memory_percent

You should adjust those settings manually to match the size of your deployment when you disable autoscaling.