In recent years, similar principles to those of DevOps in software development have been applied to machine learning projects with the goal of productionizing automated solutions. However, machine learning operations (MLOps) practices have proven to be of difficult implementation, as they often require putting together several different tools in a complex architecture. In this session, we introduce MLOps concepts and describe how we implement MLOps principles in our projects at Gucci by leveraging Databricks functionalities.

A use case of deploying a data science tool for supporting media budget allocation decisions is presented. After the development of a POC to experiment and effectively address the business problem, the code was versioned and refactored. Code reviews were executed to ensure quality and readability. Within Databricks, we rapidly moved the project to the production stage by achieving an automated solution including unit tests, environment configuration, registration and versioning of models, monitoring of model performances as well as model serving and a dashboard to visualize results.

This template is being successfully replicated for other projects and is allowing our new Data Science team to quickly bring value to different business areas of the company. These best practices and insights can be used as an example of how this can be beneficial to you or other practitioners.