Loan Prediction MLOps on AWS Platform

Production-grade loan prediction platform with MLflow model registry, DVC-backed datasets, FastAPI inference service, and monitoring/drift tooling.

What's Included

• MLOps: MLflow experiment tracking and registry, Hyperopt HPO, SHAP explainability, Fairlearn fairness checks, DVC-backed datasets (S3 remote).
• API: FastAPI service on port 8005 with /health, /prediction_api, /batch_prediction, /model_info, /list_models, and /metrics (Prometheus).
• Monitoring: Prometheus + Grafana stack (monitoring/docker-compose.yml) and Evidently-based Streamlit drift app (drift_monitoring/app_v1.py).
• Deployment: Dockerfile that syncs MLflow artifacts from s3://loanpred-mlops-20251118-120330/mlruns/, Kubernetes manifests in k8s/, and GitHub Actions CI/CD (.github/workflows/cicd.yml).

Repository Layout

Loan-Prediction_MLOps/
- .dvc/                   # DVC configuration (remote: s3://loanpred-mlops-20251118-120330)
- .github/workflows/      # CI/CD pipelines (cicd, docs-check, OIDC test)
- docs/                   # Setup, API, monitoring, CI/CD, infra guides
- drift_monitoring/       # Evidently Streamlit drift app + Dockerfile
- grafana/ , prometheus/  # Provisioning for monitoring stack
- k8s/                    # Kubernetes manifests for EKS deployment
- monitoring/             # docker-compose for Prometheus + Grafana
- prediction_model/       # Training, preprocessing, inference, config
- scripts/                # Helper scripts (DVC push, Docker build, EKS auth)
- tests/                  # API and prediction tests
- Dockerfile              # Production image (port 8005)
- datasets.dvc            # DVC tracking for datasets/ (pulled from S3)
- main.py                 # FastAPI entrypoint
- requirements*.txt       # Dependencies

Prerequisites

• Python 3.11+
• pip
• Docker + Docker Compose (for monitoring)
• AWS CLI with access to the S3 bucket (loanpred-mlops-20251118-120330 by default) for data/artifacts
• DVC with S3 support: pip install dvc[s3]

Setup

1. Clone and create a virtual environment

git clone https://github.com/your-org/Loan-Prediction_MLOps.git
cd Loan-Prediction_MLOps

# venv (Linux/Mac)
python -m venv venv && source venv/bin/activate
# venv (Windows)
python -m venv venv && venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt
# Optional: development tooling
pip install -r requirements-dev.txt

3. Configure environment variables

cp .env.example .env
# Key variables from .env.example / code defaults:
# MLFLOW_TRACKING_URI=./mlruns          # File-based MLflow tracking by default
# MODEL_STAGE=Production                # Which stage the API serves
# API_PORT=8005
# AWS_REGION=eu-west-2
# S3_BUCKET=loanpred-mlops-20251118-120330        # default in configs/DVC/Dockerfile
# DVC_REMOTE=s3://loanpred-mlops-20251118-120330
# DRIFT_S3_BUCKET=loanpred-mlops-20251118-120330

4. Pull datasets with DVC (requires AWS credentials)

aws configure                           # Ensure access to <your AWS_REGION>
dvc pull                                # Uses remote 'myremote' from .dvc/config

The default DVC remote points to s3://loanpred-mlops-20251118-120330.

5. (Optional) Download existing MLflow runs/models

aws s3 sync s3://loanpred-mlops-20251118-120330/mlruns/ ./mlruns/

Model Training

Train models and log to MLflow (experiment loan_prediction_model):

python -m prediction_model.training_pipeline

• Reads data from datasets/train.csv and datasets/test.csv (tracked by DVC).
• Logs metrics, SHAP plots, and models to ./mlruns (or the URI in MLFLOW_TRACKING_URI).
• S3 bucket used for artifacts by default: loanpred-mlops-20251118-120330 (see prediction_model/config/config.py).

Run the API

Start FastAPI (serves the latest model stage via MLflow):

python main.py
# or: uvicorn main:app --host 0.0.0.0 --port 8005

Environment notes:

• Ensure mlruns/ contains a trained model (train locally or sync from S3).
• MODEL_STAGE controls which stage to load (Production default).
• Metrics are exposed at /metrics for Prometheus scraping. Note: Inputs are parsed directly into pandas without Pydantic validation; ensure the JSON matches expected feature names and types.

Example requests

Single prediction:

curl -X POST http://localhost:8005/prediction_api \
  -H "Content-Type: application/json" \
  -d '{
    "Gender": "Male",
    "Married": "Yes",
    "Dependents": "0",
    "Education": "Graduate",
    "Self_Employed": "No",
    "ApplicantIncome": 5000,
    "CoapplicantIncome": 2000,
    "LoanAmount": 150,
    "Loan_Amount_Term": 360,
    "Credit_History": 1,
    "Property_Area": "Urban"
  }'

Batch prediction (CSV upload):

curl -X POST http://localhost:8005/batch_prediction \
  -F "file=@batch_data.csv" \
  -o predictions.csv

Model info:

curl "http://localhost:8005/model_info?stage=Production"

Monitoring and Drift

• Prometheus/Grafana:

  cd monitoring
  docker-compose up -d

  • Prometheus: http://localhost:9090
  • Grafana: http://localhost:3000 (admin/admin by default)
• The API exposes /metrics (uses prometheus_fastapi_instrumentator).
• Data drift UI:

  cd drift_monitoring
  pip install -r requirements.txt
  streamlit run app_v1.py

  Set DRIFT_S3_BUCKET if pulling batches from S3; defaults are in drift_monitoring/config.py.
• Script collect_k8s_predictions.py can collect sample predictions from a cluster endpoint for drift analysis.

Data and Artifact Management

• Datasets are tracked via DVC (datasets.dvc). Remote myremote is defined in .dvc/config:
  • URL: s3://loanpred-mlops-20251118-120330
  • Region: eu-west-2
• Common commands:

  dvc pull            # fetch datasets
  dvc add datasets/   # track updated data
  dvc push            # push data to S3

• MLflow artifacts: default bucket loanpred-mlops-20251118-120330 (Docker startup script syncs mlruns/ from there).

Docker and Deployment

• Build locally:

  scripts\build_docker_local.bat          # Windows helper
  # or
  docker build -t loan-prediction:local .

• Run container:

  docker run -p 8005:8005 \
    -e MLFLOW_TRACKING_URI="http://host.docker.internal:5000" \
    loan-prediction:local

  The image syncs mlruns/ from S3 on startup (see Dockerfile start.sh).
• Kubernetes manifests are in k8s/ (deployments, services, ServiceMonitor, namespace/auth helpers).
• GitHub Actions CI/CD (.github/workflows/cicd.yml) trains, validates metrics, pushes to ECR, and deploys to EKS using DVC data from the same S3 remote.

Testing

Tests live in tests/:

• Unit/API tests target the FastAPI app and prediction helpers.
• Integration tests are guarded by a --run-integration flag (defined in tests/conftest.py) and require MLflow models in mlruns/.
• Run:

  pytest tests/ -v                   # fast checks
  pytest tests/ -v --run-integration # requires MLflow models present in mlruns/

If you see Client.__init__() got an unexpected keyword argument 'app', reinstall the pinned HTTPX version: pip install httpx==0.24.1. If collection fails on pytest.config, update the test skip logic to use request.config.getoption("--run-integration") or the RUN_INTEGRATION flag from tests/conftest.py.

Documentation

• docs/SETUP.md: Detailed environment setup
• docs/API.md: Endpoint contract and schemas
• docs/MONITORING.md: Metrics, Grafana, Fairlearn, MLflow instructions
• docs/CI_CD_WORKFLOW.md: CI/CD pipeline breakdown
• docs/infra/*.md: ECR, EKS, Prometheus, Grafana, and S3 notes

License

MIT License. See LICENSE.