The logistic, operations and monitoring of large multinational retail enterprises required state of the art processes and algorithms. In the fashion industry the challenge doubles up with the timing component, our products simply become obsolete in months. At Inditex as a global enterprise, we serve 216 markets online and 96 markets thanks to our almost 7000 stores worldwide. We are supported by 1805 providers and almost 9000 factories. Our 8 brands sell dozens of thousands of items every year. Just compiling all the information related to these articles over time is a major Big Data problem, making use of it in a timely manner requires state of the art and custom Data Engineer and Data Science technologies. On top of that, the diversity of problems we encounter requires a large repertoire of solutions. We develop and deliver computer vision, NLP, time series analysis, clustering, regression, and classification algorithms to solve logistic, commercial, financial, sustainable, or human resources problems among others.
To deliver a machine learning solution in production, data science models are not the only key. From data collection, preprocessing, modelling, evaluation, deployment infrastructure to result monitoring and visualization, the development of the whole machine learning pipeline requires the cooperation between experts in different domains. Therefore, the machine learning solutions need to be part of a whole ecosystem that guarantees a healthy data lifecycle including scalability, monitoring, traceability, and speed. The whole data pipeline needs to be both efficient and robust, for that we bring the DevOps philosophy into machine learning via MlOps principles. MlOps focuses on reducing time to market, infrastructure as code, continuous integration, continuous delivery, and monitoring of the productized solutions.
In this talk, we will share how we developed our MLOps solution and leverage it to our Analytics projects at Inditex. How did we design our scalable, traceable, and reproducible cloud-based environment? And why is it important?
The main goal of our MLOps solution is to boost the journey between model development and machine learning production. Our MLOps framework is general enough yet specific to support the needs of the complete repertoire of use cases we encounter. Our solution includes dependency management, which maintains the same environments locally, on cloud, and in production. The machine learning pipeline supports multi-language development with parallel computing. Continuous integration and continuous delivery (CI/CD) for training, inference, and data monitoring is also several clicks away. Alerts would be triggered, and models could be retrained automatically when anomaly detected. With all these functionalities, our solution encourages data scientists and engineers to perform the best practices in all our analytics projects, which brings maintainability and scalability to our fast-evolving business.
Our obsession is helping all departments to scale, scale and scale. We leverage machine learning algorithm to automatize processes, generalize and improve decision making, and look for counterintuitive patters making the whole organization more flexible and objective. Towards this end our MlOps framework minimizes dependencies and reduces the time to market in the deployment of analytical solutions.