Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/castai-pack/skills/castai-core-workflow-a/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill castai-core-workflow-a. Copyright stays with the author (MIT).

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

Output

Return a reviewed cluster-autoscaler policy, node-template scope, capacity limits, and before/after health evidence for the nominated environment. Keep a serialized copy of the previous policy and identify the operator authorized to disable autoscaling when node churn, workload disruption, or unexpected spend breaches the guardrail.

Examples

Apply a constrained node template to a staging cluster, create one controlled unschedulable workload, and observe provisioning plus PodDisruptionBudget behavior. Promote only when capacity, lifecycle mix, and service health match the approved range; otherwise disable the policy and restore the prior template before investigating.

Resources

• Autoscaler Policies
• Node Configuration
• Terraform Node Templates

Next Steps

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