Cost Optimization

Purpose

Cloud cost optimization transforms uncontrolled spending into strategic resource allocation through the FinOps lifecycle: Inform, Optimize, and Operate. This skill provides decision frameworks for commitment-based discounts (Reserved Instances, Savings Plans), right-sizing strategies, Kubernetes cost management, and automated cost governance across multi-cloud environments.

When to Use This Skill

Invoke cost-optimization when:

• Reducing cloud spend by 15-40% through systematic optimization
• Implementing cost visibility dashboards and allocation tracking
• Establishing budget alerts and anomaly detection
• Optimizing Kubernetes resource requests and cluster efficiency
• Managing Reserved Instances, Savings Plans, or Committed Use Discounts
• Automating idle resource cleanup and right-sizing recommendations
• Setting up showback/chargeback models for internal teams
• Preventing cost overruns through CI/CD cost estimation (Infracost)
• Responding to finance team requests for cloud cost reduction

FinOps Principles

The FinOps Lifecycle

┌─────────────────────────────────────────────────────┐
│  INFORM → OPTIMIZE → OPERATE (continuous loop)      │
│    ↓         ↓           ↓                          │
│ Visibility  Action   Automation                     │
└─────────────────────────────────────────────────────┘

Inform Phase: Establish cost visibility

• Enable cost allocation tags (Owner, Project, Environment)
• Deploy real-time cost dashboards for engineering teams
• Integrate cloud billing data (AWS CUR, Azure Consumption API, GCP BigQuery)
• Set up Kubernetes cost monitoring (Kubecost, OpenCost)

Optimize Phase: Take action on cost drivers

• Purchase commitment-based discounts (40-72% savings)
• Right-size over-provisioned resources (target 60-80% utilization)
• Implement spot/preemptible instances for fault-tolerant workloads
• Clean up idle resources (unattached volumes, old snapshots)

Operate Phase: Automate and govern

• Budget alerts with cascading notifications (50%, 75%, 90%, 100%)
• Automated cleanup scripts for idle resources
• CI/CD cost estimation to prevent surprise increases
• Continuous monitoring with anomaly detection

Core FinOps Principles

1. Collaboration: Cross-functional teams (finance, engineering, operations, product)
2. Accountability: Teams own the cost of their services
3. Transparency: All costs visible and understandable to stakeholders
4. Optimization: Continuous improvement of cost efficiency

For detailed FinOps maturity models and organizational structures, see references/finops-foundations.md.

Cost Optimization Strategies

1. Commitment-Based Discounts

Reserved Instances (RIs): 40-72% discount for 1-3 year commitments

• Standard RI: Instance type locked, highest discount (60% for 3-year)
• Convertible RI: Flexible instance types, moderate discount (54% for 3-year)
• Use for: Databases (RDS, ElastiCache), stable production EC2 workloads

Savings Plans: Flexible compute commitments

• Compute Savings Plans: Applies to EC2, Fargate, Lambda (54% discount for 3-year)
• EC2 Instance Savings Plans: Tied to instance family (66% discount for 3-year)
• Use for: Workloads that change instance types or regions

GCP Committed Use Discounts (CUDs): 25-70% discount

• Resource-based CUDs: Commit to vCPU, memory, GPUs
• Spend-based CUDs: Commit to dollar amount (flexible)
• Sustained Use Discounts: Automatic 20-30% discount for sustained usage (no commitment)

Decision Framework:

Reserve when:
├─ Workload is production-critical (24/7 uptime required)
├─ Usage is predictable (stable baseline over 6+ months)
├─ Architecture is stable (unlikely to change instance types)
└─ Financial commitment acceptable (1-3 year lock-in)

Use On-Demand when:
├─ Development/testing environments
├─ Unpredictable spiky workloads
├─ Short-term projects (<6 months)
└─ Evaluating new instance types

For detailed commitment strategies and RI coverage analysis, see references/commitment-strategies.md.

2. Spot and Preemptible Instances

Discount: 70-90% off on-demand pricing (interruptible with 2-minute warning)

Use Spot For: CI/CD workers, batch jobs, ML training (with checkpointing), Kubernetes workers, data analytics Avoid Spot For: Stateful databases, real-time services, long-running jobs without checkpointing

Best Practices:

• Diversify instance types and spread across Availability Zones
• Implement graceful shutdown handlers
• Auto-fallback to on-demand when capacity unavailable
• Kubernetes: Mix 70% spot + 30% on-demand nodes with taints/tolerations

3. Right-Sizing Strategies

Target Utilization: 60-80% average (leave headroom for spikes)

Compute Right-Sizing:

• Analyze actual CPU/memory utilization over 30+ days
• Downsize instances with <40% average utilization
• Consolidate underutilized workloads
• Switch instance families (compute-optimized vs. memory-optimized)

Database Right-Sizing:

• Analyze connection pool usage (max connections vs. allocated)
• Downgrade storage IOPS if utilization <50%
• Evaluate read replica necessity (can caching replace it?)
• Consider serverless options (Aurora Serverless, Azure SQL Serverless)

Kubernetes Right-Sizing:

• Set requests = average usage (not peak)
• Set limits = 2-3x requests (allow bursting)
• Use Vertical Pod Autoscaler (VPA) for automated recommendations
• Identify pods with 0% CPU usage (candidates for consolidation)

Storage Right-Sizing:

• Delete unattached volumes (EBS, Azure Disks, GCP Persistent Disks)
• Delete old snapshots (>90 days, retention policy not required)
• Implement lifecycle policies (S3 Intelligent-Tiering, Azure Blob Lifecycle)
• Compress/deduplicate data

Right-Sizing Tools:

• AWS Compute Optimizer: ML-based EC2, Lambda, EBS recommendations
• Azure Advisor: VM rightsizing, reserved instance advice
• GCP Recommender: VM, disk, commitment recommendations
• VPA (Vertical Pod Autoscaler): Automated container resource requests

4. Kubernetes Cost Management

Resource Requests and Limits:

# Set requests = average usage (enables efficient bin-packing)
resources:
  requests:
    cpu: 500m        # 0.5 CPU cores (average usage)
    memory: 1Gi      # 1 GiB memory (average usage)
  limits:
    cpu: 1500m       # 1.5 CPU cores (3x requests, allows bursting)
    memory: 3Gi      # 3 GiB memory (3x requests)

Namespace Quotas: Prevent runaway resource consumption

• ResourceQuota: Limit total CPU/memory per namespace
• LimitRange: Default/max requests per pod
• PriorityClass: Ensure critical pods get resources

Cluster Autoscaling:

• Scale down idle nodes to reduce costs
• Scale-to-zero for dev clusters during off-hours
• Use multiple node pools (spot + on-demand mix)
• Set max node limits to prevent overspend

Cost Visibility:

• Deploy Kubecost or OpenCost for namespace-level cost tracking
• Allocate costs by labels (team, project, environment)
• Track idle cost (cluster capacity not allocated to workloads)
• Generate showback/chargeback reports

For detailed Kubernetes cost optimization patterns, see references/kubernetes-cost-optimization.md.

Cost Visibility and Monitoring

Tagging for Cost Allocation

Required Tags:

• Owner or Team - Responsible team/department
• Project or Application - Business unit or application name
• Environment - prod, staging, dev, test
• CostCenter - Finance cost center code

Enable Cost Allocation Tags:

• AWS: Activate tags in Cost Allocation Tags console
• Azure: Apply tags via Azure Policy enforcement
• GCP: Use labels on all resources, export to BigQuery

For comprehensive tagging strategies, see references/tagging-for-cost-allocation.md.

Monitoring and Dashboards

Native Cloud Tools:

• AWS Cost Explorer: Analyze spending patterns, forecast costs