stackable.md

Stackable autoscaler

This autoscaler auto-scales Stackable TrinoClusters, which are part of the Stackable Data Platform (SDP).

The SDP is an Open source data platform running on a Kubernetes cluster and offers Infrastructure as Code for Trino and related tools (such as Hive metastore, Superset or Open policy agent).

The autoscaler uses the stopped field on a Stackable TrinoCluster to turn the cluster on or off. A turned off cluster shuts down all Pods entirely (Coordinator + Workers), resulting in no computing costs at all.

Example config

Please have a look at the example config in example-configs/ha-redis-autoscaling-stackable.yaml on a working example. You need to provide at least the following two things:

Autoscaling config for cluster groups

For every cluster group that should be scaled a autoscaling configuration is needed. For cluster groups that are static (and should not be scaled) you can emit the autoscaling configuration.

trinoClusterGroups:
  s:
    maxRunningQueries: 3
    autoscaling:
      upscaleQueuedQueriesThreshold: 1
      downscaleRunningQueriesPercentageThreshold: 70
      drainIdleDurationBeforeShutdown: 60s
      minClusters:
        - timeUtc: 00:00:00 - 23:59:59
          weekdays: Mon - Son
          min: 0
    trinoClusters:
      - name: trino-s-1
        endpoint: https://trino-s-1-coordinator-default.default.svc.cluster.local:8443
        credentials: *common-credentials

In this case the cluster-group s will be started on-demand as it has a minimum cluster count of 0.

Stackable autoscaler config

The Stackable autoscaler needs to know for each TrinoCluster the Kubernetes name and namespace of the CustomResource. By having this information it can enable, disable and check Stackable TrinoClusters.

You need to add the following top-level config:

clusterAutoscaler:
  stackable:
    clusters:
      trino-s-1:
        name: trino-s-1
        namespace: default
      trino-m-1:
        name: trino-m-1
        namespace: default
      trino-m-2:
        name: trino-m-2
        namespace: default

Kubernetes requirements

trino-lb needs access to the Kubernetes cluster the Stackable Data platform is running on. The easiest way to achieve this is by running trino-lb within Kubernetes, e.g. as a Deployment with multiple replicas.

Also the ServiceAccount trino-lb is running with needs at least the following permissions to inspect and start/stop running Trino clusters:

apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: {{ .Release.Name }}
  labels:
    app.kubernetes.io/name: trino-lb
    app.kubernetes.io/instance: {{ .Release.Name }}
rules:
  - apiGroups:
      - trino.stackable.tech
    resources:
      - trinoclusters
    verbs:
      - get
      - list
      - watch
      - patch
  - apiGroups:
      - trino.stackable.tech
    resources:
      - trinoclusters/status
    verbs:
      - get