Architecture at a Glance¶

One page to understand the entire TradAI platform.

System Diagram¶

graph LR
    CLI["CLI / Users"] --> APIGW["API Gateway<br/>+ WAF + Cognito"]
    APIGW --> ALB["ALB"]

    ALB --> Backend["Backend API<br/>:8000"]
    ALB --> Strategy["Strategy Service<br/>:8003"]
    ALB --> DataCol["Data Collection<br/>:8002"]
    ALB --> MLflow["MLflow<br/>:5000"]

    Backend -->|orchestrate| Strategy
    Backend -->|data sync| DataCol
    Strategy -->|experiments| MLflow
    DataCol -->|CCXT| Exchange["Exchanges"]

    Backend -->|submit| SQS["SQS"] --> Lambda18["18 Lambda<br/>Functions"]
    Lambda18 --> SF["Step Functions<br/>Backtest + Retraining"]
    SF -->|run| ECS["ECS Tasks<br/>Freqtrade"]

    Backend --> DDB[("DynamoDB<br/>12 tables")]
    DataCol --> Arctic[("ArcticDB<br/>S3-backed")]
    MLflow --> S3[("S3<br/>Artifacts")]
    ECS --> DDB
    ECS --> MLflow

    Lambda18 --> CW["CloudWatch"] --> SNS["SNS Alerts"]

Component Quick Reference¶

Component	What It Does	Learn More
Backend API	Gateway, backtest orchestration, job management	Services, Data Flows
Strategy Service	Backtest execution, hyperopt, model promotion, A/B testing	Services, Step Functions
Data Collection	Exchange data fetching and ArcticDB storage	Services, Data Flows
MLflow	Experiment tracking, model registry, artifact storage	ML Lifecycle
Live Trading	Real-time strategy execution on ECS (partially implemented — see doc for status)	Live Trading
18 Lambda Functions	Orchestration, monitoring, ML ops, deployment	Services
Step Functions	Backtest and retraining workflows (STANDARD type)	Step Functions
VPC + Security	Network isolation, WAF, Cognito JWT auth	VPC, Security
Pulumi IaC	4-stack infrastructure (persistent, foundation, compute, edge)	Pulumi Code

Key Architecture Decisions¶

Decision	Choice	Rationale
NAT Instance over NAT Gateway	t4g.nano NAT instance	~$4/mo vs ~$32/mo; acceptable for dev/staging workloads
ECS Fargate over EC2	Fargate launch type	Zero server management; pay-per-task; simpler scaling. Note: dev/staging use EC2 consolidated mode (`CONSOLIDATED_MODE=True`) for cost savings; production uses Fargate.
Step Functions STANDARD	STANDARD over EXPRESS	Long-running backtests (up to 1 hour); built-in retry and error handling
Protocol-based DI	Python Protocols, not ABCs	Structural subtyping; no inheritance coupling; better testability
4-stack Pulumi	persistent / foundation / compute / edge	Independent lifecycle per layer; safe blast radius; parallel deployments

Reading Paths¶

If You Need	Read These (in order)
System understanding	This page then System Design then Data Flows
Deploy infrastructure	Pulumi Code then Canonical Config then Deployment
Debug an issue	State Machines then Error Handling then Observability
Understand costs	Cost Analysis
Security review	Security then VPC
ML / model lifecycle	ML Lifecycle then Step Functions
Live trading setup	Live Trading then Configuration
CI/CD and releases	CI/CD Pipeline then Deployment Pipeline
Testing approach	Testing Strategy