k8s-resource-analyzer

name	k8s-resource-analyzer
description	Analyze Kubernetes pod resource allocation vs actual usage and generate cost-saving recommendations. Use when optimizing pod requests/limits, finding cost-saving opportunities, or planning HPA adoption. Compares P95 usage metrics against allocated resources across prod-green and prod-white clusters.
version	1.0.0
category	infrastructure
tags	["kubernetes","cost-optimization","resource-management","k8s"]
metadata	{"mcp_servers":["remote-friday-mcp-server (Friday MCP for kubectl)","grafana (for Prometheus metrics)"],"setup":"See SETUP.md - requires Friday MCP and Grafana MCP registration","plan":"See PLAN.md for execution flow"}

K8s Resource Analyzer

Analyze Kubernetes deployments by comparing actual usage (P95 percentile) against allocated resources.

Overview

The skill performs four steps:

1. Fetch pod definitions from Kubernetes (Friday MCP)

   • Get unique deployments, resource requests, and limits
   • Filter to analyzable pods: running pods + completed cronjobs (excludes scaled-down/pending/failed)
   • Categorize by type: web, worker, cronjob (for clearer reporting)

2. Query Prometheus metrics (Grafana MCP)

   • CPU usage: P95 percentile over configurable time window (default: 48 hours)
   • Memory usage: P95 percentile over configurable time window
   • Configurable time ranges: 1 hour to 30 days

3. Analyze and compare

   • If no limit set → Set limit to P95 × 1.5
   • If P95 > request → Increase resources
   • If P95 < request/2 → Decrease resources (with savings tracking)
   • If variance > 30% → Enable HPA
   • If P95 > limit → Critical: Will be OOMKilled

4. Generate report

   • Organized by action:
     • 🚨 Set resource limits (missing limits)
     • ⬆️ Increase resources (under-provisioned)
     • ⬇️ Decrease resources (over-provisioned)
     • 💾 Memory optimization summary (savings vs increases)
     • 🎯 HPA candidates
     • 🔴 At-risk pods
   • Pod summary shows breakdown: "350 running (150 web + 180 workers) + 12 cronjobs"
   • Focuses on actively executing workloads

Quick Start

# Analyze default time window (48h) across both clusters
python main.py settlements

# Analyze single cluster with 24-hour window
python main.py settlements --cluster prod-green --hours 24

# Analyze 1 week (168 hours) of data
python main.py scrooge --hours 168

# Use environment variables
export K8S_NAMESPACE=settlements
export K8S_TIME_HOURS=72
python main.py

Input Parameters

The skill supports flexible configuration:

Parameter	Type	Default	Range
namespace	string	required	Any K8s namespace
--cluster / -c	string	prod-green,prod-white	prod-green, prod-white, or comma-separated
--hours / -t	int	48	1-720 hours

Environment Variables (alternative to CLI arguments):

• K8S_NAMESPACE - Kubernetes namespace (required)
• K8S_CLUSTER - Cluster to analyze (optional)
• K8S_TIME_HOURS - Time window in hours (optional)

CLI arguments take precedence over environment variables.

Prerequisites

Required MCPs:

• Friday MCP (for kubectl access)
• Grafana MCP (for Prometheus)

Setup:

1. Run: /rzp-discover:setup
2. Run: /rzp-discover:connect-mcps

See SETUP.md for detailed setup instructions.

Example Output

{
  "status": "success",
  "namespace": "settlements",
  "clusters": ["prod-green", "prod-white"],
  "pod_summary": "350 running (150 web + 180 workers) + 12 cronjobs = 362 total",
  "deployments_analyzed": 8,
  "summary": {
    "total_pods": 362,
    "total_deployments": 8
  },
  "recommendations": {
    "increase_resources": [
      {
        "deployment": "payment-processor",
        "cpu": "500m → 1500m",
        "memory": "512Mi → 1536Mi"
      }
    ],
    "decrease_resources": [
      {
        "deployment": "ledger-service",
        "cpu": "1000m → 300m",
        "memory": "1024Mi → 384Mi"
      }
    ],
    "enable_hpa": [
      {
        "deployment": "fraud-detector",
        "reason": "60% CPU variance"
      }
    ]
  }
}

Execution Flow

See PLAN.md for detailed step-by-step execution flow.

Thresholds

Metric	Condition	Action
CPU/Memory Limit	Not set	🚨 Set limit to P95 × 1.5
CPU P95	> request	⬆️ Increase to P95 × 2
CPU P95	< request/2	⬇️ Decrease to P95 × 1.5
Memory P95	> request	⬆️ Increase to P95 × 2
Memory P95	< request/2	⬇️ Decrease to P95 × 1.5
Variance	std_dev > 30%	🎯 Enable HPA
P95 > Limit	Critical risk	🔴 Will be OOMKilled

Files

File	Purpose
PLAN.md	Execution flow and detailed steps
SETUP.md	MCP setup and configuration
scripts/main.py	Orchestrator
scripts/fetch_pods.py	Query pods via Friday MCP
scripts/fetch_metrics.py	Query metrics via Grafana MCP
scripts/analyze.py	Compare and generate recommendations
scripts/report_generator.py	Format output

Notes

• Time range: Default 48 hours (customizable)
• Clusters: prod-green and prod-white
• Datasource: Primary Prometheus (UID: 6ZssswRnk)
• Output: JSON with recommendations grouped by action