Serverless Cost Engineering

Serverless promises “pay only for what you use.” In practice, serverless costs can spiral quickly if you do not understand the pricing model. A Lambda function that runs for 1 second at 1GB memory costs $0.0000167 per invocation — trivial in isolation, catastrophic at 100 million invocations per month.

---

Lambda Cost Anatomy

Lambda pricing = Requests + Duration + Memory

Requests:  $0.20 per 1 million requests
Duration:  $0.0000166667 per GB-second

Example: 10M invocations/month, 500ms average, 512MB memory
  Requests: 10M × $0.20/1M = $2.00
  Duration: 10M × 0.5s × 0.5GB × $0.0000166667 = $41.67
  Total: $43.67/month

Memory-Duration Trade-Off

Lambda CPU scales linearly with memory. More memory = more CPU = faster execution:

128MB memory:  1200ms average  →  Cost: $0.0000025
256MB memory:   800ms average  →  Cost: $0.0000033
512MB memory:   450ms average  →  Cost: $0.0000038
1024MB memory:  250ms average  →  Cost: $0.0000042

# 512MB is often the sweet spot — faster execution offsets higher memory cost

AWS Lambda Power Tuning

# Deploy the power tuning state machine
aws serverless deploy --template power-tuning.yml

# Run optimization
aws stepfunctions start-execution \
  --state-machine-arn arn:aws:states:...:power-tuning \
  --input '{"lambdaARN":"arn:aws:lambda:...:myFunction","num":50,"powerValues":[128,256,512,1024,2048]}'

---

API Gateway Costs

REST API:    $3.50 per million requests
HTTP API:    $1.00 per million requests (71% cheaper)
WebSocket:   $1.00 per million messages

Recommendation: Use HTTP API unless you need REST-specific features
  (request validation, usage plans, API keys, caching)

---

DynamoDB Cost Optimization

Capacity Modes

On-Demand:
  - $1.25 per million write requests
  - $0.25 per million read requests
  - Best for: unpredictable, spiky traffic

Provisioned:
  - $0.00065 per WCU per hour
  - $0.00013 per RCU per hour
  - Best for: predictable, steady traffic
  - 70-80% cheaper than on-demand at scale

Reserved Capacity (Provisioned):
  - 1-year or 3-year commitment
  - Additional 50-75% discount

Data Modeling for Cost

# BAD: Full table scan
response = table.scan()

# GOOD: Query with partition key
response = table.query(
    KeyConditionExpression=Key('pk').eq('USER#123')
)
# Reads only the items you need, costs proportional to data returned

---

Hidden Serverless Costs

Hidden Cost	Example	Mitigation
CloudWatch Logs	10M logs × 1KB = 10 GB ingestion = $5.00	Filter logs, reduce verbosity
Data Transfer	Cross-AZ, cross-region, internet egress	Keep services in same AZ
NAT Gateway	Lambda in VPC: $0.045/hr + $0.045/GB	Use VPC endpoints instead
Step Functions	$0.025 per 1,000 state transitions	Use Express workflows for high-volume
S3 API calls	LIST: $5 per 1M, GET: $0.40 per 1M	Cache, batch, minimize API calls

---

Cost Monitoring for Serverless

# Tag all serverless resources for cost allocation
resource_tags = {
    "Team": "order-team",
    "Service": "order-processor",
    "Environment": "production",
    "CostCenter": "CC-1234"
}

# Monitor per-function costs
# Create CloudWatch dashboard with:
#   - Invocations per function
#   - Duration per function
#   - Concurrent executions
#   - Throttles (indicates need for more capacity)

---

Anti-Patterns

Anti-Pattern	Consequence	Fix
Lambda for constant traffic	3-5x more expensive than containers	Use ECS/Fargate for sustained load
Default 128MB memory	Slow execution, often more expensive	Power-tune every function
On-demand DynamoDB at scale	Paying premium for predictable traffic	Switch to provisioned with auto-scaling
Logging everything in production	CloudWatch costs exceed Lambda costs	Structured logging with sampling
No cost tagging on functions	Cannot attribute serverless costs	Tag every function by team/service

Serverless is not always cheaper. It is always more granular — you pay exactly for usage, which can be more or less than fixed infrastructure depending on your traffic pattern.