Castai Performance Tuning

CAST AI Performance Tuning

Overview

Tune CAST AI for faster node provisioning, more responsive autoscaling, and efficient API usage. Covers headroom configuration, instance family selection, and API caching for multi-cluster dashboards.

Prerequisites

• CAST AI Phase 2 (full automation) enabled
• Understanding of workload scheduling patterns
• Access to autoscaler policy configuration

Instructions

Step 1: Optimize Node Provisioning Speed

# Configure headroom for proactive scaling (avoids waiting for pending pods)
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": {
        "enabled": true,
        "cpuPercentage": 15,
        "memoryPercentage": 15
      }
    }
  }'

Headroom pre-provisions spare capacity so pods schedule immediately instead of waiting 2-5 minutes for new nodes.

Step 2: Instance Family Optimization

# Terraform: Prefer instance families with fast launch times
resource "castai_node_template" "fast_launch" {
  cluster_id = castai_eks_cluster.this.id
  name       = "fast-launch-workers"

  constraints {
    spot                  = true
    use_spot_fallbacks    = true
    fallback_restore_rate_seconds = 300

    # Newer instance types launch faster and have better availability
    instance_families {
      include = ["m6i", "m7i", "c6i", "c7i", "r6i", "r7i"]
    }

    # Enable spot diversity for faster provisioning
    spot_diversity_price_increase_limit_percent = 25

    architectures = ["amd64"]
  }
}

Step 3: Evictor Tuning for Faster Consolidation

# Reduce empty node delay for dev/staging (faster downscale)
helm upgrade castai-evictor castai-helm/castai-evictor \
  -n castai-agent \
  --reuse-values \
  --set evictor.aggressiveMode=true \
  --set evictor.cycleInterval=120

# For production, use non-aggressive with longer intervals
# --set evictor.aggressiveMode=false
# --set evictor.cycleInterval=600

Step 4: API Performance for Multi-Cluster Dashboards

import { LRUCache } from "lru-cache";

const cache = new LRUCache<string, unknown>({ max: 100, ttl: 60_000 });

interface ClusterSummary {
  id: string;
  name: string;
  savings: number;
  savingsPercent: number;
  nodeCount: number;
  spotPercent: number;
}

async function getClusterSummary(clusterId: string): Promise<ClusterSummary> {
  const cacheKey = `summary:${clusterId}`;
  const cached = cache.get(cacheKey) as ClusterSummary | undefined;
  if (cached) return cached;

  const [cluster, savings, nodes] = await Promise.all([
    castaiGet(`/v1/kubernetes/external-clusters/${clusterId}`),
    castaiGet(`/v1/kubernetes/clusters/${clusterId}/savings`),
    castaiGet(`/v1/kubernetes/external-clusters/${clusterId}/nodes`),
  ]);

  const spotNodes = nodes.items.filter(
    (n: { lifecycle: string }) => n.lifecycle === "spot"
  ).length;

  const summary: ClusterSummary = {
    id: clusterId,
    name: cluster.name,
    savings: savings.monthlySavings,
    savingsPercent: savings.savingsPercentage,
    nodeCount: nodes.items.length,
    spotPercent: nodes.items.length > 0
      ? (spotNodes / nodes.items.length) * 100
      : 0,
  };

  cache.set(cacheKey, summary);
  return summary;
}

// Aggregate across all clusters
async function getDashboardData(
  clusterIds: string[]
): Promise<ClusterSummary[]> {
  return Promise.all(clusterIds.map(getClusterSummary));
}

Step 5: Workload Autoscaler Tuning

# Faster resource adjustment with shorter cooldown
# (use with caution in production)
metadata:
  annotations:
    autoscaling.cast.ai/cpu-headroom: "10"     # Lower headroom = tighter fit
    autoscaling.cast.ai/memory-headroom: "15"
    autoscaling.cast.ai/apply-type: "immediate" # Apply without waiting

Performance Benchmarks

Error Handling

Output

Publish a bounded tuning record containing the measured baseline, target SLO, proposed setting, expected cost and disruption impact, rollout scope, and rollback trigger. Treat estimates as hypotheses: production success requires observed node, workload, and cost telemetry over an agreed window.

Examples

In staging, shorten a workload-autoscaler cooldown for one non-critical service, with a PodDisruptionBudget and alerting already in place. Compare provision time, eviction count, p95 latency, and spend against the saved baseline; if disruption or error rates rise, restore the prior setting before testing another variable.

Resources

• Autoscaler Settings
• Workload Autoscaler Annotations
• Node Configuration

Next Steps

For cost optimization strategies, see castai-cost-tuning.