Advanced Elasticsearch node scheduling

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

These features can be combined together, to deploy a production-grade Elasticsearch cluster.

Define Elasticsearch nodes roles

edit

You can configure Elasticsearch nodes with one or multiple roles. This allows you to describe an Elasticsearch cluster with 3 dedicated master nodes, for example:

apiVersion: elasticsearch.k8s.elastic.co/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  nodeSets:
  # 3 dedicated master nodes
  - name: master
    count: 3
    config:
      node.master: true
      node.data: false
      node.ingest: false
      cluster.remote.connect: false
  # 3 ingest-data nodes
  - name: ingest-data
    count: 3
    config:
      node.master: false
      node.data: true
      node.ingest: true

Affinity options

edit

You can setup 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/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  nodesSets:
  - 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 that behaviour, set an empty affinity object:

apiVersion: elasticsearch.k8s.elastic.co/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  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/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  nodeSets:
  - name: default
    count: 3
    podTemplate:
      spec:
        affinity:
          podAntiAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
            - podAffinityTerm:
                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/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  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/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  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 from being scheduled on Kubernetes hosts tagged with environment: e2e or environment: production. It favors nodes tagged with diskType: ssd.

Availability zone awareness

edit

By combining Elasticsearch shard allocation awareness with Kubernetes node affinity, you can setup an availability zone-aware Elasticsearch cluster:

apiVersion: elasticsearch.k8s.elastic.co/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  nodeSets:
  - name: zone-a
    count: 1
    config:
      node.attr.zone: europe-west3-a
      cluster.routing.allocation.awareness.attributes: zone
    podTemplate:
      spec:
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: failure-domain.beta.kubernetes.io/zone
                  operator: In
                  values:
                  - europe-west3-a
  - name: zone-b
    count: 1
    config:
      node.attr.zone: europe-west3-b
      cluster.routing.allocation.awareness.attributes: zone
    podTemplate:
      spec:
        affinity:
          nodeAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: failure-domain.beta.kubernetes.io/zone
                  operator: In
                  values:
                  - europe-west3-b

This example relies on:

Kubernetes nodes in each zone being labeled accordingly. failure-domain.beta.kubernetes.io/zone is standard, but any label can be used.
node affinity for each group of nodes set to match the Kubernetes nodes' zone.
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 setup an Elasticsearch cluster with hot-warm topology:

apiVersion: elasticsearch.k8s.elastic.co/v1beta1
kind: Elasticsearch
metadata:
  name: quickstart
spec:
  version: 8.16.0
  nodeSets:
  # hot nodes, with high CPU and fast IO
  - name: hot
    count: 3
    config:
      node.attr.data: hot
    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.attr.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). Here we restrict pods to be scheduled on Kubernetes nodes labeled with beta.kubernetes.io/instance-type: highio (to adapt to your Kubernetes nodes' labels).
the second group has the data attribute set to warm. It is intended to run on hosts with larger but maybe slower storage. Pods are only able to 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, setup Index Lifecycle Management policies on your indices, optimizing for hot-warm architectures.

« Nodes orchestration Create automated snapshots »