This documentation contains work-in-progress information for future Elastic Stack and Cloud releases. Use the version selector to view supported release docs. It also contains some Elastic Cloud serverless information. Check out our serverless docs for more details.

« Nodes orchestration Create automated snapshots »

› › ›

Advanced Elasticsearch node scheduling

edit

Advanced Elasticsearch node scheduling

edit

Elastic Cloud on Kubernetes (ECK) offers full control over Elasticsearch cluster nodes scheduling by combining Elasticsearch configuration with Kubernetes scheduling options:

You can combine these features to deploy a production-grade Elasticsearch cluster.

Define Elasticsearch nodes roles

edit

You can configure Elasticsearch nodes with one or multiple roles.

You can use YAML anchors to declare the configuration change once and reuse it across all the node sets.

This allows you to describe an Elasticsearch cluster with 3 dedicated master nodes, for example:

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  # 3 dedicated master nodes
  - name: master
    count: 3
    config:
      node.roles: ["master"]
  # 3 ingest-data nodes
  - name: ingest-data
    count: 3
    config:
      node.roles: ["data", "ingest"]

Affinity options

edit

You can set up various affinity and anti-affinity options through the podTemplate section of the Elasticsearch resource specification.

A single Elasticsearch node per Kubernetes host (default)

edit

To avoid scheduling several Elasticsearch nodes from the same cluster on the same host, use a podAntiAffinity rule based on the hostname and the cluster name label:

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  - name: default
    count: 3
    podTemplate:
      spec:
        affinity:
          podAntiAffinity:
            preferredDuringSchedulingIgnoredDuringExecution:
            - weight: 100
              podAffinityTerm:
                labelSelector:
                  matchLabels:
                    elasticsearch.k8s.elastic.co/cluster-name: quickstart
                topologyKey: kubernetes.io/hostname

This is ECK default behaviour if you don’t specify any affinity option. To explicitly disable the default behaviour, set an empty affinity object:

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  - name: default
    count: 3
    podTemplate:
      spec:
        affinity: {}

The default affinity is using preferredDuringSchedulingIgnoredDuringExecution, which acts as best effort and won’t prevent an Elasticsearch node from being scheduled on a host if there are no other hosts available. Scheduling a 4-nodes Elasticsearch cluster on a 3-host Kubernetes cluster would then successfully schedule 2 Elasticsearch nodes on the same host. To enforce a strict single node per host, specify requiredDuringSchedulingIgnoredDuringExecution instead:

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  - name: default
    count: 3
    podTemplate:
      spec:
        affinity:
          podAntiAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchLabels:
                  elasticsearch.k8s.elastic.co/cluster-name: quickstart
              topologyKey: kubernetes.io/hostname

Local Persistent Volume constraints

edit

By default, volumes can be bound to a Pod before the Pod gets scheduled to a particular Kubernetes node. This can be a problem if the PersistentVolume can only be accessed from a particular host or set of hosts. Local persistent volumes are a good example: they are accessible from a single host. If the Pod gets scheduled to a different host based on any affinity or anti-affinity rule, the volume may not be available.

To solve this problem, you can set the Volume Binding Mode of the StorageClass you are using. Make sure volumeBindingMode: WaitForFirstConsumer is set, especially if you are using local persistent volumes.

kind: StorageClass
apiVersion: storage.k8s.io/v1
metadata:
  name: local-storage
provisioner: kubernetes.io/no-provisioner
volumeBindingMode: WaitForFirstConsumer

Node affinity

edit

To restrict the scheduling to a particular set of Kubernetes nodes based on labels, use a NodeSelector. The following example schedules Elasticsearch Pods on Kubernetes nodes tagged with both labels diskType: ssd and environment: production.

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  - name: default
    count: 3
    podTemplate:
      spec:
        nodeSelector:
          diskType: ssd
          environment: production

You can achieve the same (and more) with node affinity:

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  - name: default
    count: 3
    podTemplate:
      spec:
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: environment
                  operator: In
                  values:
                  - e2e
                  - production
            preferredDuringSchedulingIgnoredDuringExecution:
              - weight: 1
                preference:
                  matchExpressions:
                  - key: diskType
                    operator: In
                    values:
                    - ssd

This example restricts Elasticsearch nodes so they are only scheduled on Kubernetes hosts tagged with environment: e2e or environment: production. It favors nodes tagged with diskType: ssd.

Topology spread constraints and availability zone awareness

edit

Starting with ECK 2.0 the operator can make Kubernetes Node labels available as Pod annotations. It can be used to make information, such as logical failure domains, available in a running Pod. Combined with Elasticsearch shard allocation awareness and Kubernetes topology spread constraints, you can create an availability zone-aware Elasticsearch cluster.

Exposing Kubernetes node topology labels in Pods

edit

First, ensure that the operator’s flag exposed-node-labels contains the list of the Kubernetes node labels that should be exposed in the Elasticsearch Pods. If you are using the provided installation manifest, or the Helm chart, then this flag is already preset with two wildcard patterns for well-known node labels that describe Kubernetes cluster topology, like topology.kubernetes.io/.* and failure-domain.beta.kubernetes.io/.*.
On the Elasticsearch resources set the eck.k8s.elastic.co/downward-node-labels annotations with the list of the Kubernetes node labels that should be copied as Pod annotations.
Use the Kubernetes downward API in the podTemplate to make those annotations available as environment variables in Elasticsearch Pods.

Refer to the next section or to the Elasticsearch sample resource in the ECK source repository for a complete example.

Using node topology labels, Kubernetes topology spread constraints, and Elasticsearch shard allocation awareness

edit

The following example demonstrates how to use the topology.kubernetes.io/zone node labels to spread a NodeSet across the availability zones of a Kubernetes cluster.

Note that by default ECK creates a k8s_node_name attribute with the name of the Kubernetes node running the Pod, and configures Elasticsearch to use this attribute. This ensures that Elasticsearch allocates primary and replica shards to Pods running on different Kubernetes nodes and never to Pods that are scheduled onto the same Kubernetes node. To preserve this behavior while making Elasticsearch aware of the availability zone, include the k8s_node_name attribute in the comma-separated cluster.routing.allocation.awareness.attributes list.

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  annotations:
    eck.k8s.elastic.co/downward-node-labels: "topology.kubernetes.io/zone"
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  - name: default
    count: 3
    config:
      node.attr.zone: ${ZONE}
      cluster.routing.allocation.awareness.attributes: k8s_node_name,zone
    podTemplate:
      spec:
        containers:
        - name: elasticsearch
          env:
          - name: ZONE
            valueFrom:
              fieldRef:
                fieldPath: metadata.annotations['topology.kubernetes.io/zone']
        topologySpreadConstraints:
          - maxSkew: 1
            topologyKey: topology.kubernetes.io/zone
            whenUnsatisfiable: DoNotSchedule
            labelSelector:
              matchLabels:
                elasticsearch.k8s.elastic.co/cluster-name: quickstart
                elasticsearch.k8s.elastic.co/statefulset-name: quickstart-es-default

This example relies on:

Kubernetes nodes in each zone being labeled accordingly. topology.kubernetes.io/zone is standard, but any label can be used.
Pod topology spread constraints to spread the Pods across availability zones in the Kubernetes cluster.
Elasticsearch configured to allocate shards based on node attributes. Here we specified node.attr.zone, but any attribute name can be used. node.attr.rack_id is another common example.

Hot-warm topologies

edit

By combining Elasticsearch shard allocation awareness with Kubernetes node affinity, you can set up an Elasticsearch cluster with hot-warm topology:

apiVersion: elasticsearch.k8s.elastic.co/v1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.17.0
  nodeSets:
  # hot nodes, with high CPU and fast IO
  - name: hot
    count: 3
    config:
      node.roles: ["data_hot", "ingest", "master"]
    podTemplate:
      spec:
        containers:
        - name: elasticsearch
          resources:
            limits:
              memory: 16Gi
              cpu: 4
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: beta.kubernetes.io/instance-type
                  operator: In
                  values:
                  - highio
    volumeClaimTemplates:
    - metadata:
        name: elasticsearch-data
      spec:
        accessModes:
        - ReadWriteOnce
        resources:
          requests:
            storage: 1Ti
        storageClassName: local-storage
  # warm nodes, with high storage
  - name: warm
    count: 3
    config:
      node.roles: ["data_warm"]
    podTemplate:
      spec:
        containers:
        - name: elasticsearch
          resources:
            limits:
              memory: 16Gi
              cpu: 2
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: beta.kubernetes.io/instance-type
                  operator: In
                  values:
                  - highstorage
    volumeClaimTemplates:
    - metadata:
        name: elasticsearch-data
      spec:
        accessModes:
        - ReadWriteOnce
        resources:
          requests:
            storage: 10Ti
        storageClassName: local-storage

In this example, we configure two groups of Elasticsearch nodes:

The first group has the data attribute set to hot. It is intended to run on hosts with high CPU resources and fast IO (SSD). Pods can only be scheduled on Kubernetes nodes labeled with beta.kubernetes.io/instance-type: highio (to adapt to the labels of your Kubernetes nodes).
The second group has the data attribute set to warm. It is intended to run on hosts with larger but maybe slower storage. Pods can only be scheduled on nodes labeled with beta.kubernetes.io/instance-type: highstorage.

This example uses Local Persistent Volumes for both groups, but can be adapted to use high-performance volumes for hot Elasticsearch nodes and high-storage volumes for warm Elasticsearch nodes.

Finally, set up Index Lifecycle Management policies on your indices, optimizing for hot-warm architectures.

« Nodes orchestration Create automated snapshots »