"Why is our AWS bill so high? We're barely using any compute."

Answer: NAT Gateway data processing charges.

What was happening:

  • All external API calls went through NAT Gateway
  • $0.045 per GB processed
  • Our ML pipeline downloaded 500GB of training data daily
  • NAT Gateway bill: $675/day
  • Monthly: $20,250 just for data processing

The architecture problem:

  • Private subnets for "security"
  • Everything routed through NAT Gateway
  • S3 downloads going internet → NAT → private subnet

The fix:

  • S3 VPC endpoint for training data (free!)
  • Private subnets for internal services
  • NAT only for services that actually need internet
  • VPC endpoints for AWS services (DynamoDB, SQS, etc.)

Result: $20,250/month → $1,200/month.

Lesson: VPC endpoints exist. Use them. They're free for S3 (Gateway endpoint) and cheap for others.


← Zurück zu Erfahrungsberichte