The Architecture-Cost Connection
Most engineers think about cost as an operational concern - something the finance team worries about after the system is built. This is backwards.
The decisions that drive the largest cloud costs are made during system design: how services communicate, where data lives, how often it moves, and what gets cached. By the time the bill arrives, the architecture is already set.
The Most Expensive Architectural Mistakes
1. Cross-AZ data transfer in hot paths
AWS charges for data transfer between Availability Zones. A microservice architecture where services in different AZs call each other on every request can generate significant data transfer costs that are invisible during design.
Mitigation: Co-locate services that communicate frequently. Use AZ-aware load balancing.
2. NAT Gateway for internal traffic
Traffic from private subnets to AWS services (S3, DynamoDB, SQS) routes through NAT Gateway by default - at $0.045/GB. For high-throughput systems, this is a significant cost.
Mitigation: Use VPC endpoints for AWS services accessed from private subnets.
3. Synchronous fan-out to downstream services
A single API request that triggers synchronous calls to five downstream services multiplies your compute and data transfer costs by five. Under load, this compounds.
Mitigation: Async event-driven patterns where latency allows. Cache aggressively at the aggregation layer.
4. Storing everything in RDS when DynamoDB fits
RDS costs are dominated by instance size, not usage. A multi-AZ RDS instance runs 24/7 regardless of load. For access patterns that fit DynamoDB's model, the cost difference at scale is an order of magnitude.
Cost as an Architectural Fitness Function
Add cost to your architectural fitness functions:
# cost-fitness.yml
thresholds:
monthly_compute: $2000
monthly_data_transfer: $500
monthly_storage: $300
alerts:
- metric: aws.billing.estimated_charges
threshold: 80%
action: notify_architect
When cost is a first-class architectural concern, it gets the same rigour as latency and availability.