Skip to content

Rossmann Sales Forecasting

Production MLOps System

This repository accompanies the ML/AI System Design book, providing a hands-on implementation of the concepts covered in the course.

This project demonstrates end-to-end MLOps best practices for traditional machine learning, using the Kaggle Rossmann Store Sales competition as a real-world example. The reason for this example is because many students and individuals new to MLOps are often exposed to Kaggle examples—this is a great way to learn how to build production ML systems using familiar datasets.

Overview

The Business Problem

This project is based on the Kaggle Rossmann Store Sales competition, which challenges participants to forecast daily sales for Rossmann drugstore chains across Europe. Rossmann operates over 3,000 stores in 7 European countries, and accurate sales forecasting is critical for inventory management, staffing decisions, and financial planning.

The competition dataset includes:

  • Historical sales data for 1,115 stores from 2013-2015
  • Store metadata including type, assortment, competition distance
  • Promotional information for regular and extended promotions
  • External factors like holidays, school closures, and seasonality

The forecasting challenge requires predicting sales for a 6-week period, with performance measured by RMSPE (Root Mean Square Percentage Error).

The Hypothetical Scenario

Imagine you work as a data scientist at Rossmann. You've successfully developed a forecasting model that achieves strong performance (RMSPE \< 0.10) on historical data. Now, your director has tasked you with productionizing this system to support ongoing business operations.

The requirements:

  • Automated daily/weekly updates: New sales data arrives regularly and must be processed automatically
  • Data quality guarantees: Invalid data must be caught before it corrupts the system
  • Model retraining: The model should retrain automatically as new data becomes available
  • Production deployment: Forecasts must be accessible via API and dashboard for business stakeholders
  • Monitoring and alerts: System must detect data drift and performance degradation
  • Reproducibility: All data, models, and predictions must be versioned and auditable

This repository provides an exemplar approach to productionizing a Kaggle-style model, assuming:

  • New actual sales data arrives on a daily or weekly basis
  • Historical data grows continuously, improving model performance over time
  • The system must run reliably with minimal human intervention
  • Quality and reproducibility are more important than speed

What This Project Demonstrates

The Rossmann Sales Forecasting project is a production-ready ML system that showcases modern MLOps practices:

  • 📊 DataOps: Data versioning (DVC), validation (Great Expectations), reproducible pipelines
  • 🤖 ModelOps: Experiment tracking (MLflow), model registry, hyperparameter tuning
  • 🚀 Deployment: REST API (FastAPI), interactive dashboard (Streamlit), containerization (Docker)
  • 📈 Monitoring: Data drift detection (Evidently), model performance tracking
  • ✅ CI/CD: Automated testing (pytest), GitHub Actions workflows

This is not just a Kaggle competition submission—it's a blueprint for building production ML systems that handle the messy realities of real-world data and operational constraints.

Key Features

Production-Grade Architecture

  • Modular code structure with clear separation of concerns
  • Comprehensive test suite with >80% coverage
  • Type hints and documentation throughout
  • Modern Python packaging with uv

MLOps Infrastructure

  • Experiment Tracking: All runs logged to MLflow
  • Data Validation: Automated quality checks with Great Expectations
  • Version Control: Data and models versioned with DVC
  • Containerization: Docker for reproducible deployments
  • Monitoring: Drift detection and performance tracking

  • Getting Started

    Install and set up the project in minutes

    Quick Start

  • DataOps

    Learn data processing, validation, and versioning

    DataOps Guide

  • ModelOps

    Train, track, and register models with MLflow

    ModelOps Guide

  • Deployment

    Deploy models via API and dashboard (Coming Soon)

Technology Stack

Core ML

  • Python 3.10+ with modern packaging (uv, pyproject.toml)
  • Pandas for data manipulation
  • Scikit-learn for preprocessing and metrics
  • LightGBM, XGBoost, CatBoost for gradient boosting

MLOps Tools

  • MLflow - Experiment tracking and model registry
  • DVC - Data and pipeline versioning
  • Great Expectations - Data validation
  • Evidently - Drift detection

Deployment

  • FastAPI - REST API for predictions
  • Streamlit - Interactive dashboard
  • Docker - Containerization
  • GitHub Actions - CI/CD automation

Development

  • pytest - Testing framework
  • black & ruff - Code formatting and linting
  • mypy - Type checking
  • MkDocs Material - Documentation