AWS Cost Optimization: An SRE's Playbook

The Bill That Woke Me Up

Three years ago, I got an AWS bill for $2,800. For a side project. That was the moment I stopped treating cloud costs as someone else's problem.

Since then, I've cut that bill to $1,800/month while increasing reliability and capacity. This isn't theoretical — these are the exact techniques I use, with real numbers.

1. Right-Sizing: The Low-Hanging Fruit

Most teams over-provision because they guessed during setup. AWS's own data shows 40% of EC2 instances are under-utilized.

# Check your actual utilization
aws cloudwatch get-metric-statistics \
  --namespace AWS/EC2 \
  --metric-name CPUUtilization \
  --dimensions Name=InstanceId,Value=i-1234567890 \
  --start-time $(date -u -d '14 days ago' +%Y-%m-%dT%H:%M:%S) \
  --end-time $(date -u +%Y-%m-%dT%H:%M:%S) \
  --period 86400 \
  --statistics Average

Rule of thumb: If your average CPU is below 20% and memory below 40% for 14 days, you can probably drop one instance size. I saved $180/month just by moving a t3.xlarge to t3.large.

2. Reserved Instances and Savings Plans

If you're running anything stable (databases, core services, CI/CD), you're leaving money on the table without commitments.

• 1-year No Upfront RI: ~30% savings on baseline compute
• Compute Savings Plans: ~25% savings with flexibility across instance families
• S3 Intelligent-Tiering: Automatically moves objects between access tiers

I run Savings Plans on my baseline (3x m6i.large) and use Spot for burst workloads. This alone saves me ~$400/month.

3. Spot Instances for Fault-Tolerant Workloads

Spot instances are 60-90% cheaper than On-Demand. The catch: they can be interrupted with 2 minutes notice. Use them for:

• CI/CD build runners
• Data processing jobs
• Stateless web servers behind an ALB
• Kubernetes worker nodes (with proper drain handling)

# Spot Fleet request for CI runners
{
  "SpotFleetRequestConfig": {
    "AllocationStrategy": "capacityOptimized",
    "TargetCapacity": 2,
    "LaunchSpecifications": [
      {
        "InstanceType": "c5.xlarge",
        "ImageId": "ami-12345678",
        "KeyName": "ci-key",
        "SecurityGroups": [{"GroupId": "sg-ci"}]
      }
    ]
  }
}

I use Spot for all my CI/CD pipelines. My average build cost dropped from $0.17/hour to $0.04/hour.

4. Storage: Where Costs Hide

Storage costs sneak up on you. EBS volumes, S3 buckets, snapshots — they accumulate silently.

• Delete unattached EBS volumes: Check monthly, automate the cleanup
• Snapshot lifecycle policies: Keep 7 daily, 4 weekly, 12 monthly — not forever
• S3 lifecycle rules: Move to Glacier after 90 days, expire after 365
• EBS gp3 > gp2: Same price, 20% better baseline IOPS, tunable

# Find unattached EBS volumes (often $50-200/month wasted)
aws ec2 describe-volumes \
  --filters Name=status,Values=available \
  --query 'Volumes[*].{ID:VolumeId,Size:Size,Type:VolumeType}'

5. Network Costs: The Silent Killer

Data transfer is where AWS makes its real money. Inter-AZ traffic, NAT Gateway, and cross-region replication add up fast.

• NAT Gateway vs NAT Instance: A t3.micro NAT instance costs $7/month. A NAT Gateway costs $32/month + data processing fees. For low-traffic environments, use the instance.
• VPC Endpoints: S3 and DynamoDB Gateway endpoints are free. They keep traffic off the NAT.
• Consolidate services in same AZ: Inter-AZ traffic is $0.01/GB each way.

6. Automated Cleanup

Manual cleanup doesn't scale. Set up automation:

# Lambda: Delete old EBS snapshots (older than 90 days)
import boto3
from datetime import datetime, timedelta

def lambda_handler(event, context):
    ec2 = boto3.client('ec2')
    cutoff = datetime.now() - timedelta(days=90)
    
    snapshots = ec2.describe_snapshots(OwnerIds=['self'])
    for snap in snapshots['Snapshots']:
        if snap['StartTime'].replace(tzinfo=None) < cutoff:
            ec2.delete_snapshot(SnapshotId=snap['SnapshotId'])
            print(f"Deleted: {snap['SnapshotId']}")

7. Environment Tiering

Not every environment needs production-grade infrastructure:

• Production: RIs, multi-AZ, reserved capacity
• Staging: Spot + On-Demand mix, single-AZ acceptable
• Dev/Test: Scheduled shutdown (Lambda to stop instances at 8 PM), Spot only, smallest viable sizes

My dev environment auto-stops at night and on weekends. That's a 65% reduction in non-production costs.

The Results

Before optimization:
  EC2:        $1,200/mo
  RDS:          $450/mo
  S3/EBS:       $320/mo
  Data Xfer:    $280/mo
  NAT/Network:  $150/mo
  Other:        $400/mo
  Total:      $2,800/mo

After optimization:
  EC2 (Spot+RI): $680/mo
  RDS (RI):       $310/mo
  S3 (Lifecycle): $180/mo
  Data Xfer:      $120/mo
  NAT (VPC EP):    $45/mo
  Other:          $265/mo
  Total:        $1,800/mo

Savings: $1,000/month (36%)

Where to Start

1. Week 1: Enable AWS Cost Explorer. Identify your top 5 cost drivers.
2. Week 2: Right-size your instances. Delete unattached volumes.
3. Week 3: Set up Spot for non-production workloads.
4. Week 4: Purchase Savings Plans for your baseline.
5. Ongoing: Monthly cost review. Automate cleanup.

Cost optimization isn't a one-time project. It's an ongoing discipline. Start with the biggest wins (right-sizing + Spot), then layer on the refinements.

Questions about your AWS bill? Get in touch.