scale-sniff

Detect wasteful or under-provisioned Kubernetes autoscaling in on-prem environments

scale-sniff analyzes CPU utilization metrics against Pod Counts over time to identify inefficient scaling patterns in on-prem Kubernetes clusters. Pinpoint over-provisioned deployments, under-utilized resources, and misconfigured HPA policies to optimize costs and performance.

Why Use scale-sniff?

• Reduce costs by identifying over-provisioned resources
• Improve performance by detecting under-provisioned workloads
• Optimize HPA configurations with data-driven insights
• On-prem focused – designed for environments without cloud auto-scaling benefits

Contributions welcome! Please open a GitHub issue for:

• 🐛 Bug reports
• 💡 Feature requests
• ❓ Questions & discussions
• 📝 General feedback

🔍 How It Works

• Discovers Services → finds corresponding Deployments via pod label selectors
• Maps App to Pods
  • Service → selects Pods via spec.selector
  • Pods ← owned by Deployment ← scaled by HPA
• Fetches Metrics from Prometheus:
  • CPU: from cAdvisor (built into kubelet)
  • Pod Count: from kube-state-metrics
• Analyzes Efficiency:
  • Compares actual CPU to HPA target with google/gemma-2-2b-it model
  • Flags over-provisioning (high pods, low CPU) or under-provisioning (high CPU, low pods)

📊 Requirements

Component	Purpose
kube-state-metrics	Exposes kube_pod_status_ready for pod count
cAdvisor (default in kubelet)	Provides container_cpu_usage_seconds_total for CPU usage
Prometheus	Scrapes and stores metrics
HPA	Enables autoscaling analysis
HuggingFace token	Analyze step requires HuggingFace token Visit Hugging Face Token Documentation for more details.

🚀 Quick Start

go build -o scale-sniff ./cmd/cli

# help
Usage: ./scale-sniff [options]

Options:
  -config-path string
        Configuration file path (config.yaml should be under that!) (default "./internal/config")
  -duration string
        Analysis window (default "24h")
  -h    Show help
  -help
        Show help
  -hf-model string
        HuggingFace model override (default "google/gemma-2-2b-it")
  -hf-token string
        HuggingFace API token
  -hf-url string
        HuggingFace API URL (default "https://router.huggingface.co/v1/chat/completions")
  -namespace string
        Kubernetes namespace (default "default")
  -prometheus-base-url string
        Prometheus base URL
  -range-vector string
        Rate range vector (default "5m")
  -step string
        Sample resolution (default "5m")

# Analyze a namespace
./scale-sniff analyze --namespace prod

# Custom time window
./scale-sniff analyze --namespace dev --duration 15m --step 30s

🎬 Demo Setup

Prerequisites

k3d (or any local Kubernetes cluster)
kubectl
go (for building the CLI)

Apply in order under demo directory:

Create local cluster (optional if using existing cluster)
k3d cluster create demo --agents 1

kubectl apply -f nginx-deployment.yaml

kubectl apply -f prometheus-rbac.yaml

kubectl apply -f prometheus-config.yaml

kubectl apply -f prometheus-deployment.yaml

kubectl apply -f ksm-rbac.yaml   

kubectl apply -f kube-state-metrics.yaml

kubectl get pods
kubectl get services

# Or use validation script 
chmod +x validate.sh
./validate.sh

Create HPA

kubectl autoscale deployment nginx-demo --cpu-percent=10 --min=1 --max=10

Forward prometheus port

 kubectl port-forward service/prometheus-service 9090:9090

Export HuggingFace Token

 export HUGGINGFACE_TOKEN=<your-token>

Run scale-sniff

Run scale-sniff under project directory

 ./scale-sniff 
⏳ nginx-service: fetching
⏳ prometheus-service: fetching
⏳ nginx-service: analyzing

📊 Final Report:
❓ prometheus-service: Inconclusive — No HPA configured — autoscaling not enabled
✅ nginx-service: Efficient

clean up

k3d cluster delete demo