MLFlow Demo Project

This repository demonstrates a complete MLOps pipeline using MLflow, DagsHub, DVC, and Evidently for comprehensive model lifecycle management, data versioning, and experiment tracking.

Features

• ** MLflow Integration**: Experiment tracking, model registry, and artifact logging
• ** DagsHub Pipeline**: Complete MLOps collaboration platform with visual pipelines
• ** DVC Data Versioning**: Version-controlled datasets and model artifacts
• ** Drift Detection**: Evidently-based data drift monitoring
• ** Automated Pipeline**: End-to-end reproducible ML workflows

Project Structure

MLflow_demo/
│
├── data/
│   ├── data_loader.py
│   ├── raw/                    # Original dataset (DVC tracked)
│   ├── processed/              # Cleaned data (pipeline output)
│   └── drift_baseline/         # Drift detection reports
│
├── notebooks/
│   ├── 01_data_cleaning.ipynb
│   ├── 02_drift.ipynb
│   └── 03_model_training.ipynb
│
├── src/
│   ├── data_preprocessing.py   # Data cleaning pipeline
│   ├── drift_detection.py      # Evidently drift detection
│   ├── train.py               # Model training with MLflow
│   ├── evaluate.py            # Model evaluation
│   └── pipeline.py            # Complete end-to-end pipeline
│
├── dvc.yaml                   # DVC pipeline configuration
├── dvc.lock                   # Pipeline lock file
├── metrics.json               # Pipeline metrics output
├── requirements.txt
└── README.md

Setup

Prerequisites

• Python 3.11+
• Git
• DVC
• MLflow
• DagsHub account

Installation

1. Clone the repository:

   git clone https://github.com/yahiaehab10/MLFlow_demo.git
   cd MLFlow_demo
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Configure DagsHub authentication (for data push):

   # Get your token from: https://dagshub.com/user/settings/tokens
   dvc remote modify origin password <your-dagshub-token>
   

Usage

Run Complete Pipeline

Option 1: DVC Pipeline (Recommended)

dvc repro  # Runs the complete reproducible pipeline

Option 2: Direct Python Execution

python -m src.pipeline  # Runs with MLflow logging to DagsHub

Individual Components

• Data Processing: python -m src.data_preprocessing
• Model Training: python -m src.train
• Drift Detection: python -m src.drift_detection

DVC Commands

• View Pipeline: dvc dag
• Check Pipeline Status: dvc status
• Push Data to DagsHub: dvc push
• Pull Data from DagsHub: dvc pull

MLflow & DagsHub Integration

Experiment Tracking

All experiments are automatically tracked and logged to:

• DagsHub MLflow: https://dagshub.com/yahiaehab10/MLFlow_demo.mlflow
• Local MLflow UI: Run mlflow ui and visit http://localhost:5000

Model Registry & Stage Management

• Models are automatically registered as IrisRandomForest in MLflow Model Registry
• Support for model stage transitions (Staging, Production)
• Model versioning and lineage tracking
• Artifact storage (drift reports, model files, metrics)

Logged Metrics & Parameters

• Parameters: Data paths, model hyperparameters, random seeds
• Metrics: Accuracy, precision, recall
• Artifacts:
  • Drift baseline HTML reports
  • Model files (pickle format)
  • Analysis images (when available)

DagsHub Features

Data Pipeline Visualization

• Visual representation of your ML pipeline
• Dependency tracking between stages
• Real-time status monitoring

Collaboration

• Version-controlled datasets with DVC
• Experiment comparison and analysis
• Team collaboration on ML projects
• Model performance monitoring

Data Drift Detection

The pipeline includes comprehensive drift detection using Evidently:

• Baseline Generation: Creates reference profiles from training data
• Drift Monitoring: Compares new data against baseline
• HTML Reports: Interactive drift analysis reports
• Automated Alerts: Integration with MLflow for drift tracking

Pipeline Workflow

1. Data Loading: Load raw iris dataset
2. Data Cleaning: Preprocess and clean data
3. Drift Detection: Generate baseline and detect drift
4. Model Training: Train RandomForest with MLflow tracking
5. Model Registration: Register model in MLflow registry
6. Metrics Logging: Save performance metrics
7. Artifact Upload: Store reports and models in DagsHub

Data Versioning with DVC

• Raw Data: data/raw/iris.csv - Version controlled with DVC
• Processed Data: data/processed/iris_clean.csv - Pipeline output
• Models: Stored in MLflow model registry
• Reports: data/drift_baseline/ - Drift analysis artifacts

Configuration

Environment Variables

The pipeline uses the following configuration:

• DagsHub Repository: yahiaehab10/MLFlow_demo
• MLflow Tracking URI: https://dagshub.com/yahiaehab10/MLFlow_demo.mlflow
• Model Name: IrisRandomForest

DVC Configuration

# dvc.yaml
stages:
  full_pipeline:
    cmd: python -m src.pipeline
    deps:
      - data/raw/iris.csv
      - src/pipeline.py
      - src/data_preprocessing.py
      - src/train.py
      - src/drift_detection.py
    outs:
      - data/processed/iris_clean.csv
      - data/drift_baseline/iris_drift_baseline.html
    metrics:
      - metrics.json

Troubleshooting

Common Issues

1. DVC Push Authentication Error:

   dvc remote modify origin password <your-dagshub-token>
   

2. MLflow Tracking URI Error:

   • Ensure you have internet connectivity
   • Verify DagsHub repository access

3. Pipeline Dependencies:

   pip install -r requirements.txt
   

Documentation

• MLflow: MLflow Documentation
• DagsHub: DagsHub Tutorial
• DVC: DVC Documentation
• Evidently: Evidently Documentation

License

This project is open source and available under the MIT License.