TradAI Final Architecture - Cost Analysis¶

NAT Gateway (2 AZs):
├─ Hourly charge: $0.045/hr × 730 hrs × 2 = $65.70
├─ Data processing: $0.045/GB × ~100 GB = $4.50
└─ Total: ~$70/month

NAT Instance (t4g.nano):
├─ Instance: $0.0042/hr × 730 hrs = $3.07
├─ EIP: $0 (attached to instance)
├─ Data: Included
└─ Total: ~$3/month + some overhead = $6.13

SAVINGS: ~$64/month (91% reduction)

Trade-offs:
- Lower bandwidth (5 Gbps vs 45 Gbps) - acceptable for our workload
- Self-managed (patching, monitoring) - minimal effort with ASG
- Single point of failure - mitigated with ASG health checks

VPC Interface Endpoints (currently deployed -- see 03-VPC-NETWORKING.md):
├─ ECR API:          ~$7.30/mo/AZ × 2 AZs = $14.60
├─ ECR DKR:          ~$7.30/mo/AZ × 2 AZs = $14.60
├─ STS:              ~$7.30/mo/AZ × 2 AZs = $14.60
├─ Secrets Manager:  ~$7.30/mo/AZ × 2 AZs = $14.60
├─ CloudWatch Logs:  ~$7.30/mo/AZ × 2 AZs = $14.60
├─ SSM:              ~$7.30/mo/AZ × 2 AZs = $14.60
├─ SSM Messages:     ~$7.30/mo/AZ × 2 AZs = $14.60
├─ EC2 Messages:     ~$7.30/mo/AZ × 2 AZs = $14.60
├─ SQS:              ~$7.30/mo/AZ × 2 AZs = $14.60
└─ Total: ~$65.70/month (9 interface endpoints)

NOTE: This optimization was NOT applied. The 9 interface endpoints
ARE deployed (per 03-VPC-NETWORKING.md Section 6.2). The cost tables
above assume endpoint removal ($0 in Networking optimized column),
but actual spend includes ~$65.70/month for these endpoints.

Potential savings if endpoints are removed:
├─ Route all traffic through NAT Instance instead
├─ Additional NAT traffic: ~10 GB/month = $0.45/month
└─ SAVINGS: ~$65/month

Trade-offs of removal:
- Slightly higher latency for AWS API calls (~10ms)
- Traffic goes through internet (still encrypted via TLS)
- Acceptable for non-high-frequency operations
- SSM Session Manager would require NAT for access

RDS Multi-AZ:
├─ Primary: $18/month
├─ Standby: $18/month
└─ Total: $36/month

RDS Single-AZ:
└─ Total: $18/month

SAVINGS: $18/month (50% reduction)

Trade-offs:
- No automatic failover
- 15-30 minute recovery on failure
- Acceptable for dev/staging environments

Production recommendation:
- Keep Multi-AZ for production
- Use Single-AZ for dev/staging

Fargate On-Demand (Strategy Tasks):
├─ 50 runs × 30 min × $0.04/hr = $10/month
└─ Total: ~$10/month

Fargate Spot:
├─ 70% discount on Fargate pricing
├─ 50 runs × 30 min × $0.012/hr = $3/month
└─ Total: ~$3/month

SAVINGS: $7/month (70% reduction)

Trade-offs:
- Tasks may be interrupted (2-minute warning)
- Mitigation: Checkpoint progress to S3
- Acceptable for batch processing (backtests)

CloudWatch Logs pricing:
├─ Ingestion: $0.50/GB (one-time, on ingest)
├─ Storage: $0.03/GB-month (NOT per-day — this is a monthly rate)

30-day retention:
├─ 10 GB × $0.50/GB ingestion = $5.00
├─ 10 GB × $0.03/GB-month storage = $0.30
└─ Total: ~$5.30/month

7-day retention:
├─ 10 GB × $0.50/GB ingestion = $5.00
├─ ~2.3 GB avg stored × $0.03/GB-month = $0.07
└─ Total: ~$5.07/month

SAVINGS: $0.23/month (negligible — ingestion dominates CW Logs cost)

Note: Reducing retention saves very little money. To meaningfully cut
CW Logs costs, reduce log VOLUME (structured logging, sampling) or
export to S3 ($0.023/GB-month) for long-term storage.

Trade-offs:
- Shorter debugging window
- Mitigation: Export important logs to S3 for long-term storage

Without lifecycle:
├─ Temp files accumulate
├─ 100 GB × $0.023/GB = $2.30
└─ Growing over time

With lifecycle:
├─ Temp files deleted after 7 days
├─ Results archived to Glacier after 30 days
├─ ~40 GB average = $1.00
└─ Stable cost

SAVINGS: $1.30/month + prevents cost growth

RDS On-Demand:
└─ db.t4g.micro: $18/month

RDS Reserved (1-year, no upfront):
└─ db.t4g.micro: $12/month

SAVINGS: $6/month (33% reduction)

Requirements:
- 1-year commitment
- Wait 1 month to verify stable usage

Current architecture supports 50-200 backtests/month at ~$128-137/month

At 500+ backtests/month:
- Consider provisioned Step Functions
- Consider reserved Fargate capacity
- Estimated cost: ~$165/month

At 1000+ backtests/month:
- Consider dedicated EC2 instances for backtesting
- Consider Kubernetes (EKS) for orchestration
- Estimated cost: ~$315/month

Budget Configuration:
├─ Monthly Budget: $135
├─ Alert 1: 50% ($68) - Email notification
├─ Alert 2: 80% ($108) - Email + Slack notification
├─ Alert 3: 100% ($135) - Email + Slack + PagerDuty
└─ Forecast Alert: 110% projected - Email notification

Required Tags:
├─ Application: tradai
├─ Environment: production | staging | dev
├─ Service: backend-api | data-collection | mlflow | strategy
├─ Owner: team-name
└─ CostCenter: trading-platform

Monthly by Service:
- Filter: Application = tradai
- Group by: Service tag
- Time: Last 3 months

Daily Trend:
- Filter: Application = tradai
- Group by: Day
- Time: Last 30 days

Anomaly Detection:
- Enable AWS Cost Anomaly Detection
- Threshold: 20% above normal
- Alert: Email + Slack