ML Platform POC - Quick Start

This is a local Proof of Concept demonstrating the ML platform architecture before production deployment.

🚀 Quick Start (5 minutes)

# 0. Convert CSV to Parquet (one-time setup)
python scripts/convert_csv_to_parquet.py

# 1. Start infrastructure
docker-compose up -d

# 2. Install dependencies
# may require python -m ensurepip --upgrade if pip is not available
pip install -r requirements-poc.txt

# 3. Train model
python poc/train_fraud_model.py

# 4. Populate feature store
python poc/populate_features.py

# 5. Start model server
python poc/serve_model.py

# 6. Test (in another terminal)
python poc/test_predictions.py

📊 What's Included

Core Components

• MLflow - Experiment tracking & model registry (http://localhost:5000)
• Cassandra - Online feature store (localhost:9042)
• Feast - Feature store framework (optional, production-grade)
• Prometheus - Metrics & monitoring (http://localhost:9090)
• Fraud Detection Model - XGBoost classifier with <5ms latency

Model Serving Options

• FastAPI + Direct Cassandra - Simple serving (http://localhost:8000)

  • Best for: Quick development, testing
  • Setup time: 30 seconds
  • Feature store: Direct Cassandra access

• FastAPI + Feast - Production-grade serving (http://localhost:8000)

  • Best for: Feature versioning, consistency guarantees
  • Setup time: 5 minutes
  • Feature store: Feast with Cassandra backend

• Seldon Core - Kubernetes-based serving (http://localhost:8001)

  • Best for: Production parity, auto-scaling, A/B testing
  • Setup time: 10 minutes
  • Requires: Local Kubernetes cluster

📁 Project Structure

ml_deployment/
├── README.md                    # This file
├── docker-compose.yml          # Infrastructure setup
├── requirements-poc.txt        # Python dependencies
├── docs/                       # Documentation
│   ├── POC_LOCAL_SETUP.md      # Detailed setup guide
│   ├── ml_platform_design.md   # Production architecture design
│   ├── PARQUET_MIGRATION.md    # Parquet format guide
│   ├── FEAST_SETUP.md          # Feast feature store guide
│   └── SELDON_SETUP.md         # Kubernetes deployment guide
├── data/
│   ├── credit_card_fraud_10k.parquet  # Fraud detection data (Parquet)
│   ├── Twitter_Data.parquet           # Sentiment analysis data
│   └── ...                            # Other datasets
├── poc/
│   ├── train_fraud_model.py    # Model training
│   ├── populate_features.py    # Feature store setup
│   ├── serve_model.py          # Model serving (FastAPI)
│   ├── test_predictions.py     # Functional testing
│   └── load_test.py            # Performance testing
├── feature_repo/               # Feast feature definitions
├── kubernetes/                 # Kubernetes deployment files
├── scripts/                    # Utility scripts
└── mlflow_data/                # MLflow artifacts (created on first run)

🧪 Testing

# Functional tests
python poc/test_predictions.py

# Load test (100 RPS for 30 seconds)
python poc/load_test.py --rps 100 --duration 30

# High load test (1000 RPS)
python poc/load_test.py --rps 1000 --duration 10

📈 Expected Performance

• Latency: < 5ms (P99)
• Throughput: 300+ RPS per CPU core
• Model Accuracy: AUC > 0.94, F1 > 0.89
• Feature Fetch: ~2-3ms from Cassandra

🔍 Monitoring

• MLflow UI: http://localhost:5000 - View experiments and models
• API Docs: http://localhost:8000/docs - Interactive API documentation
• Prometheus: http://localhost:9090 - Query metrics
• Metrics Endpoint: http://localhost:8000/metrics - Raw metrics

📚 Documentation

• Quick Start: This file (README.md)
• Detailed Setup: See docs/POC_LOCAL_SETUP.md
• Production Design: See docs/ml_platform_design.md
• Parquet Format: See docs/PARQUET_MIGRATION.md (data optimization)
• Feast Integration: See docs/FEAST_SETUP.md (feature store framework)
• Seldon Core Setup: See docs/SELDON_SETUP.md (optional, for K8s deployment)

🛠️ Troubleshooting

Server won't start:

# Check if ports are available
lsof -i :5000  # MLflow
lsof -i :9042  # Cassandra
lsof -i :8000  # Model server
lsof -i :9090  # Prometheus

Model not found:

# Retrain the model
python poc/train_fraud_model.py

Cassandra connection error:

# Restart Cassandra
docker-compose restart cassandra

# Check Cassandra status
docker-compose ps cassandra

# Note: Cassandra takes 30-60 seconds to start
docker-compose logs -f cassandra

🧹 Cleanup

# Stop services
docker-compose down

# Remove all data
docker-compose down -v
rm -rf mlflow_data/

➡️ Next Steps

After validating the POC:

Phase 1: Basic Validation

1. ✅ Confirm latency < 5ms with FastAPI
2. ✅ Verify model accuracy meets requirements
3. ✅ Test with your own data
4. 📝 Document lessons learned

Phase 2: Advanced (Optional)

5. 🔧 Deploy with Seldon Core on local Kubernetes
6. 📊 Compare FastAPI vs Seldon Core performance
7. 🧪 Test auto-scaling and A/B testing features

Phase 3: Production Planning

8. 🚀 Plan production migration (see docs/ml_platform_design.md)

🤝 Support

For questions or issues, refer to:

• docs/POC_LOCAL_SETUP.md for detailed instructions
• docs/ml_platform_design.md for architecture decisions