Skills Development Configure CAST AI Autoscaler Policies

Configure CAST AI Autoscaler Policies

v20260423
castai-core-workflow-a
This workflow guides users through configuring CAST AI's autoscaler policies to achieve optimal cost management and resource utilization within a Kubernetes cluster. It covers enabling spot instances, setting node downscaler/evictor rules, defining cluster limits, and creating specialized node templates (e.g., for GPU or spot workers) using Terraform. Use this when implementing Phase 2 automation or optimizing cloud spend.
Get Skill
492 downloads
Overview

CAST AI Core Workflow: Autoscaler & Policies

Overview

Primary workflow for CAST AI: configure autoscaler policies to optimize cluster costs. Covers enabling spot instances, configuring the node downscaler and evictor, setting cluster CPU/memory limits, and creating node templates for workload-specific requirements.

Prerequisites

  • Completed castai-install-auth with Phase 2 (cluster controller + evictor)
  • CASTAI_API_KEY and CASTAI_CLUSTER_ID set
  • Cluster in "ready" status

Instructions

Step 1: Read Current Policies

curl -s -H "X-API-Key: ${CASTAI_API_KEY}" \
  "https://api.cast.ai/v1/kubernetes/clusters/${CASTAI_CLUSTER_ID}/policies" \
  | jq .

Step 2: Enable Cost-Optimized Autoscaling

curl -X PUT -H "X-API-Key: ${CASTAI_API_KEY}" \
  -H "Content-Type: application/json" \
  "https://api.cast.ai/v1/kubernetes/clusters/${CASTAI_CLUSTER_ID}/policies" \
  -d '{
    "enabled": true,
    "unschedulablePods": {
      "enabled": true,
      "headroom": {
        "cpuPercentage": 10,
        "memoryPercentage": 10,
        "enabled": true
      }
    },
    "nodeDownscaler": {
      "enabled": true,
      "emptyNodes": {
        "enabled": true,
        "delaySeconds": 180
      }
    },
    "spotInstances": {
      "enabled": true,
      "clouds": ["aws"],
      "spotDiversityEnabled": true,
      "spotDiversityPriceIncreaseLimitPercent": 20
    },
    "clusterLimits": {
      "enabled": true,
      "cpu": {
        "minCores": 4,
        "maxCores": 100
      }
    }
  }'

Step 3: Configure Node Templates via Terraform

resource "castai_node_template" "spot_workers" {
  cluster_id = castai_eks_cluster.this.id
  name       = "spot-workers"
  is_default = false
  is_enabled = true

  constraints {
    min_cpu               = 2
    max_cpu               = 16
    min_memory            = 4096
    max_memory            = 65536
    spot                  = true
    use_spot_fallbacks    = true
    fallback_restore_rate_seconds = 600

    instance_families {
      include = ["m5", "m6i", "c5", "c6i", "r5", "r6i"]
    }

    architectures = ["amd64"]
  }

  custom_labels = {
    "workload-type" = "batch"
  }
}

resource "castai_node_template" "gpu_ondemand" {
  cluster_id = castai_eks_cluster.this.id
  name       = "gpu-ondemand"
  is_default = false
  is_enabled = true

  constraints {
    spot                  = false
    gpu_manufacturers     = ["NVIDIA"]

    instance_families {
      include = ["p3", "p4d", "g4dn", "g5"]
    }
  }

  custom_labels = {
    "workload-type" = "gpu"
  }
}

Step 4: Verify Autoscaler is Working

# Check if the autoscaler is processing nodes
curl -s -H "X-API-Key: ${CASTAI_API_KEY}" \
  "https://api.cast.ai/v1/kubernetes/external-clusters/${CASTAI_CLUSTER_ID}/nodes" \
  | jq '[.items[] | {name, instanceType, lifecycle, castaiManaged: .castaiManaged}]
        | group_by(.lifecycle)
        | map({lifecycle: .[0].lifecycle, count: length})'

# Expected: mix of spot and on-demand nodes

Error Handling

ErrorCauseSolution
Policy update returns 400Invalid policy JSONValidate with jq before sending
Nodes not scalingPolicy not enabledVerify .enabled: true in policy
Spot instances not usedProvider not configuredAdd cloud provider to spotInstances.clouds
Evictor too aggressiveLow delay thresholdIncrease emptyNodes.delaySeconds
Cluster limit hitmaxCores too lowIncrease clusterLimits.cpu.maxCores

Resources

Next Steps

For workload-level autoscaling, see castai-core-workflow-b.

Info
Category Development
Name castai-core-workflow-a
Version v20260423
Size 4.42KB
Updated At 2026-04-28
Language