mlops-serving-platform

Plataforma de serving de modelo ML (Isolation Forest) via FastAPI, containerizada com Docker e deployada em Kubernetes. Portfólio de MLOps end-to-end.

Stack

Camada	Tecnologia
API	FastAPI + uvicorn
Modelo	scikit-learn Isolation Forest
Containerização	Docker multi-stage
Dev	Docker Compose
Produção	Kubernetes (EKS)
CI/CD	GitHub Actions + AWS ECR
Observabilidade	Prometheus

Rodar localmente

Pré-requisitos

• Docker + Docker Compose
• Python 3.11

1. Treinar o modelo

pip install -r app/requirements.txt
python scripts/train_model.py

2. Subir a stack

cd compose
docker-compose up --build

Acesse:

• API: http://localhost/health
• Prometheus: http://localhost:9090

3. Testar o endpoint /predict

curl -X POST http://localhost/predict \
  -H "Content-Type: application/json" \
  -d '{"features": [0.1, -0.2, 0.3, 0.0]}'

Response:

{"score": -0.142, "is_anomaly": false}

Deploy em produção (EKS)

Pré-requisitos

• Cluster EKS configurado
• ECR repository criado
• Secrets no GitHub: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, EKS_CLUSTER_NAME, ECR_REGISTRY

Deploy automático

Push para main dispara o pipeline completo: testes → build → push ECR → deploy EKS.

Deploy manual

export ECR_REGISTRY=<sua-conta>.dkr.ecr.us-east-1.amazonaws.com
export IMAGE_TAG=latest

python scripts/train_model.py
docker build -f docker/Dockerfile -t $ECR_REGISTRY/mlops-serving:$IMAGE_TAG .
docker push $ECR_REGISTRY/mlops-serving:$IMAGE_TAG

sed -i "s|\${ECR_REGISTRY}|$ECR_REGISTRY|g" k8s/deployment.yaml
sed -i "s|\${IMAGE_TAG}|$IMAGE_TAG|g" k8s/deployment.yaml
kubectl apply -f k8s/

Testes

python scripts/train_model.py
pytest tests/ -v

Arquitetura

Ver ARCHITECTURE.md