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

CAST AI Core Workflow: Workload Autoscaler

Overview

CAST AI Workload Autoscaler right-sizes pod resource requests based on actual usage, reducing over-provisioning without manual VPA tuning. This skill covers enabling the workload autoscaler, configuring scaling policies per workload, and using annotations for fine-grained control.

Prerequisites

• Completed castai-core-workflow-a (cluster-level policies)
• CAST AI agent v1.60+ installed
• Workload Autoscaler enabled in CAST AI console

Instructions

Step 1: Install Workload Autoscaler Components

helm upgrade --install castai-workload-autoscaler \
  castai-helm/castai-workload-autoscaler \
  -n castai-agent \
  --set castai.apiKey="${CASTAI_API_KEY}" \
  --set castai.clusterID="${CASTAI_CLUSTER_ID}"

Step 2: Query Workload Recommendations

# Get resource recommendations for a specific workload
curl -s -H "X-API-Key: ${CASTAI_API_KEY}" \
  "https://api.cast.ai/v1/workload-autoscaling/clusters/${CASTAI_CLUSTER_ID}/workloads" \
  | jq '.items[] | {
    name: .workloadName,
    namespace: .namespace,
    currentCpu: .currentCpuRequest,
    recommendedCpu: .recommendedCpuRequest,
    currentMemory: .currentMemoryRequest,
    recommendedMemory: .recommendedMemoryRequest,
    savingsPercent: .estimatedSavingsPercent
  }'

Step 3: Configure Per-Workload Policies via Annotations

# Add annotations to deployments for CAST AI workload autoscaler
apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-api
  annotations:
    # Enable workload autoscaling
    autoscaling.cast.ai/enabled: "true"
    # CPU configuration
    autoscaling.cast.ai/cpu-min: "100m"
    autoscaling.cast.ai/cpu-max: "4000m"
    autoscaling.cast.ai/cpu-headroom: "15"
    # Memory configuration
    autoscaling.cast.ai/memory-min: "128Mi"
    autoscaling.cast.ai/memory-max: "8Gi"
    autoscaling.cast.ai/memory-headroom: "20"
    # Apply changes automatically vs recommendation-only
    autoscaling.cast.ai/apply-type: "immediate"
spec:
  template:
    spec:
      containers:
        - name: api
          resources:
            requests:
              cpu: "500m"      # Will be auto-adjusted by CAST AI
              memory: "512Mi"  # Will be auto-adjusted by CAST AI

Step 4: Create a Scaling Policy via API

curl -X POST -H "X-API-Key: ${CASTAI_API_KEY}" \
  -H "Content-Type: application/json" \
  "https://api.cast.ai/v1/workload-autoscaling/clusters/${CASTAI_CLUSTER_ID}/policies" \
  -d '{
    "name": "cost-optimized",
    "applyType": "IMMEDIATE",
    "management": {
      "cpu": {
        "function": "QUANTILE",
        "args": { "quantile": 0.95 },
        "overhead": 0.15,
        "min": 50,
        "max": 8000
      },
      "memory": {
        "function": "MAX",
        "overhead": 0.20,
        "min": 64,
        "max": 16384
      }
    },
    "antiShrink": {
      "enabled": true,
      "cooldownSeconds": 300
    }
  }'

Step 5: Monitor Workload Scaling Events

# Check scaling events
kubectl get events -n default --field-selector reason=CastAIWorkloadAutoscaled

# View current vs recommended via API
curl -s -H "X-API-Key: ${CASTAI_API_KEY}" \
  "https://api.cast.ai/v1/workload-autoscaling/clusters/${CASTAI_CLUSTER_ID}/workloads/${WORKLOAD_ID}" \
  | jq '.scalingEvents[-5:]'

Error Handling

Output

Produce an approved workload-autoscaler policy, the observed request/limit baseline, selected guardrails, change ticket, and before/after workload health evidence. Keep the policy scoped to the named workload and retain the prior configuration so it can be restored if latency, errors, or eviction behavior regresses.

Examples

Apply a conservative policy to one staging deployment with a 15 percent memory overhead and a five-minute anti-shrink cooldown. Observe a controlled demand change, compare p95 latency and restart counts to the baseline, then promote only after service owners approve the evidence; revert the annotation if the workload OOMs or violates its disruption budget.

Resources

• Workload Autoscaler Overview
• Annotations Reference
• Scaling Policies

Next Steps

For troubleshooting CAST AI errors, see castai-common-errors.