ModelServe

MLOps with Cloud Season 2 — Capstone Exam

A production-ready fraud detection inference service with MLflow model registry, Feast feature store, Prometheus monitoring, and Grafana dashboards.

Prerequisites

• Docker and Docker Compose
• Python 3.12+
• AWS account (for Pulumi infrastructure)
• Redis, PostgreSQL (provided via Docker Compose)

Quick Start (Local Development)

# 1. Start all services
docker-compose up -d

# 2. Wait for services to be healthy (check with:)
docker ps

# 3. Train the model
source mlops_venv/bin/activate
python training/train.py

# 4. Materialize features to Redis (from host or container)
# From host:
python -c "
from feast import FeatureStore
from datetime import datetime, timezone
fs = FeatureStore(repo_path='./feast_repo')
fs.materialize(start_date=datetime(2024,1,1,tzinfo=timezone.utc), end_date=datetime.now(timezone.utc))
"
# Or from container:
docker exec modelserve-fastapi python -c "
from feast import FeatureStore
from datetime import datetime, timezone
fs = FeatureStore(repo_path='/app/feast_repo')
fs.materialize(start_date=datetime(2024,1,1,tzinfo=timezone.utc), end_date=datetime.now(timezone.utc))
"

# 5. Test the API
curl http://localhost:8000/health
curl -X POST http://localhost:8000/predict -H "Content-Type: application/json" -d '{"entity_id": 340187018810220}'

Services

Service	Port	Description
FastAPI	8000	Inference API
MLflow	5000	Model registry & tracking
Prometheus	9090	Metrics collection
Grafana	3001	Dashboards
PostgreSQL	5432	MLflow backend store
Redis	6379	Feast online store

REST Endpoints

Method	Path	Description
GET	/health	Health check with model version
POST	/predict	Make fraud prediction
GET	/predict/{entity_id}?explain=true	Prediction with feature values
GET	/metrics	Prometheus metrics

Environment Variables

See .env.example for all required variables:

Variable	Description	Default
POSTGRES_USER	PostgreSQL username	mlflow
POSTGRES_PASSWORD	PostgreSQL password	mlflow
MLFLOW_TRACKING_URI	MLflow server URI	http://localhost:5000
MODEL_NAME	Model name in registry	FraudDetector
MODEL_STAGE	Model stage to load	Production
REDIS_HOST	Redis host	redis
REDIS_PORT	Redis port	6379

GitHub Secrets

Secret	Purpose
AWS_ACCESS_KEY_ID	AWS credentials for S3
AWS_SECRET_ACCESS_KEY	AWS credentials for S3
AWS_REGION	AWS region

Engineering Documentation

• docs/README.md — documentation index
• docs/ARCHITECTURE.md — system overview, Mermaid architecture diagrams, five ADRs, CI/CD, runbook, known limitations
• docs/diagrams/ — how to export diagrams from the Mermaid in ARCHITECTURE.md
• guide/issues.md — known issues and troubleshooting

Dataset

Credit Card Transactions Fraud Detection — Simulated credit card transactions generated using Sparkov. Use fraudTrain.csv (~1.3M rows, 22 features). Entity key: cc_num.

---

MLOps with Cloud Season 2 — Capstone: ModelServe | Poridhi.io