Cloud Cost Optimization#

The average company wastes 30-35% of cloud spend on idle or oversized resources. Cloud cost optimization is not about cutting corners — it is about paying only for what you actually use.

The Problem#

Month 1: Spin up dev cluster "temporarily"     → $800/mo
Month 3: Nobody remembers who owns it           → still running
Month 6: Finance asks "why is AWS $47K/mo?"     → panic
Month 7: Team finds 23 idle load balancers      → $4,200/mo wasted

Cloud makes it easy to provision. It makes it hard to de-provision.

Right-Sizing#

Most instances run at 10-20% CPU utilization. Right-sizing means matching instance size to actual usage.

Before:

Production API: c5.4xlarge (16 vCPU, 32 GB)
  Avg CPU: 12%
  Avg Memory: 18%
  Monthly cost: $490

After right-sizing:

Production API: c5.xlarge (4 vCPU, 8 GB)
  Avg CPU: 48%
  Avg Memory: 72%
  Monthly cost: $122
  Savings: 75%

How to identify candidates:

1. Pull 14-day CPU/memory metrics from CloudWatch
2. Flag instances with peak utilization below 40%
3. Recommend one size down (or two if peak is below 20%)
4. Test in staging first
5. Apply during maintenance window

Reserved Instances and Savings Plans#

For stable workloads, commit to 1-3 years for massive discounts:

Strategy:

Baseline load (always running)     → Reserved Instances / Savings Plans
Variable load (scales with traffic) → On-Demand + Auto-Scaling
Batch / fault-tolerant jobs        → Spot Instances
Dev/test environments              → Spot + scheduled shutdown

Spot Instances#

Spare cloud capacity at 60-90% discount. The catch: they can be reclaimed with 2 minutes notice.

Good for:

→ CI/CD build runners
→ Batch data processing
→ Stateless web workers behind a load balancer
→ Machine learning training jobs (with checkpointing)
→ Test environments

Bad for:

→ Databases
→ Stateful services without checkpointing
→ Single-instance critical services

Spot interruption strategy:

1. Use multiple instance types (diversify capacity pools)
2. Use multiple AZs
3. Set up interruption handler: save state → drain connections → terminate
4. Mix spot + on-demand in ASG (e.g., 70% spot, 30% on-demand)

Auto-Scaling#

Scale resources to match demand — pay for peak only when peak happens.

Target tracking:

Auto-Scaling Policy:
  Metric: Average CPU utilization
  Target: 60%
  Min instances: 2
  Max instances: 20
  Cooldown: 300 seconds

Traffic spike → CPU rises above 60% → add instances
Traffic drops → CPU falls below 60% → remove instances

Scheduled scaling for predictable patterns:

Weekdays 9 AM:  scale to 10 instances (business hours)
Weekdays 7 PM:  scale to 3 instances  (evening)
Weekends:       scale to 2 instances   (minimal traffic)

Savings: ~55% vs running 10 instances 24/7

Idle Resource Detection#

The silent budget killer: resources nobody uses but everybody pays for.

Common culprits:

→ Unattached EBS volumes              ($0.10/GB/mo, adds up fast)
→ Idle load balancers                  ($16-22/mo each, zero traffic)
→ Stopped instances with attached EBS  (compute free, storage not)
→ Old snapshots                        ($0.05/GB/mo, forgotten)
→ Unused Elastic IPs                   ($3.60/mo each)
→ Orphaned NAT Gateways               ($32/mo + data processing)
→ Dev/staging clusters left running    (hundreds/mo)

Automated cleanup script pattern:

Daily scan:
  1. List all EBS volumes where state = "available" (unattached)
  2. List all ELBs with zero requests in 14 days
  3. List all instances with avg CPU below 2% for 7 days
  4. List all snapshots older than 90 days
  5. Generate report → send to Slack channel
  6. Auto-tag with "idle-candidate" and "cleanup-date"
  7. Delete after 14-day grace period if unclaimed

FinOps: The Cultural Practice#

FinOps is not a tool — it is a practice where engineering, finance, and business collaborate on cloud spending.

Three phases:

Inform  → Visibility into who spends what
          → Dashboards, cost allocation, showback reports

Optimize → Right-size, reserve, spot, auto-scale
           → Engineering-driven changes with financial context

Operate → Continuous governance
          → Budgets, alerts, anomaly detection, policy enforcement

Key FinOps metrics:

Unit economics:   cost per transaction, cost per user, cost per request
Coverage ratio:   % of spend covered by commitments (target: 70-80%)
Waste ratio:      % of spend on idle resources (target: below 5%)
Forecast accuracy: predicted vs actual spend (target: within 5%)

Tagging Strategy#

Tags are the foundation of cost visibility. Without tags, you cannot allocate costs.

Mandatory tags for every resource:

team:        "platform"
environment: "production"
service:     "payment-api"
owner:       "jane@company.com"
cost-center: "engineering-42"
created-by:  "terraform"

Enforcement:

1. AWS SCP / Azure Policy → deny resource creation without required tags
2. CI check → Terraform plan must include all mandatory tags
3. Weekly scan → find untagged resources → auto-notify owner
4. Dashboard → show % of spend that is tagged (target: 95%+)

Cost Allocation#

Map cloud spend to teams, products, and features:

Total AWS bill: $47,000/mo

By team:
  Platform:    $18,000 (38%)
  Data:        $14,000 (30%)
  Product:     $10,000 (21%)
  Untagged:     $5,000 (11%) ← fix this

By environment:
  Production:  $28,000 (60%)
  Staging:      $9,000 (19%)
  Dev:          $5,000 (11%)
  Untagged:     $5,000 (11%)

Showback vs chargeback:

Showback:   "Team X, your cloud spend was $18K this month" (informational)
Chargeback: "Team X, $18K is deducted from your budget" (accountability)

Start with showback. Move to chargeback once tagging is mature.

Tools#

Infracost in CI#

Estimate cost impact before merging infrastructure changes:

PR opens → Terraform plan → Infracost diff
                              ↓
                     "This change adds $240/mo"
                              ↓
                     Reviewer sees cost impact in PR comment
                              ↓
                     Approve or request optimization

Kubecost#

Real-time Kubernetes cost allocation by namespace, deployment, and label:

Kubecost Dashboard:
  namespace/production:  $12,400/mo
  namespace/staging:      $3,200/mo
  namespace/monitoring:   $1,800/mo

  Over-provisioned pods:  47 (requesting 4x actual usage)
  Recommended savings:    $4,100/mo

Track your infrastructure costs at codelit.io — generate interactive architecture diagrams with resource and cost annotations.

Summary#

1. Right-size instances — match capacity to actual utilization
2. Reserve stable workloads — 1-year Savings Plans for 30-40% off
3. Spot for fault-tolerant jobs — 60-90% savings with interruption handling
4. Auto-scale everything — pay for peak only during peak
5. Kill idle resources — automated scans, grace periods, then cleanup
6. Tag religiously — enforce mandatory tags via policy
7. FinOps culture — inform, optimize, operate continuously

---

362 articles on system design at codelit.io/blog.