feat: Add offline MLflow support with local artifact fallback

[Sottim]

Offline MLflow Support

Overview

This PR adds resilient model loading with automatic fallback to local artifacts when MLflow server is unavailable, ensuring production reliability.

Files Modified

• src/model/inference.py - Add fallback loading logic
• src/api/__init__.py - Update error handling
• docker-compose.yml - Add volume mounts

Features

• Graceful Degradation: Auto-fallback when MLflow unavailable
• Zero Downtime: Service continues without MLflow server
• Offline Support: Works in disconnected environments
• Multiple Path Search: Checks several locations for artifacts
• Full Compatibility: Existing MLflow workflow unchanged

Technical Implementation

Before:

def load(self):
    model_uri = f"runs:/{self.run_id}/{self.model_flavor}"
    self.model = mlflow.lightgbm.load_model(model_uri)
    # Fails if MLflow server is down

After:

def load(self):
    try:
        self._load_from_mlflow()  # Try MLflow first
    except Exception:
        self._load_from_local()    # Fallback to local
    # Always works now

Local Artifact Structure

notebooks/mlartifacts/
└── {EXPERIMENT_ID}/
    └── {RUN_ID}/
        └── artifacts/
            └── {model_flavor}/
                └── model.pkl

Testing

Test with MLflow:

docker compose up -d --build mlflow prod
curl http://localhost:8000/predict
# Loads from MLflow server

Test without MLflow:

docker compose stop mlflow
docker compose restart prod
curl http://localhost:8000/predict
#  Loads from local artifacts