How MLOps Infrastructures Are Built at Drizly

To further stir up the discussions among the community, I decided to summarize my findings at Drizly to demonstrate how the company built the ML infrastructure stack. In this article, I will explain the uniqueness of our situations, as well as showcase the rationales behind infrastructure decisions. I hope that our lessons will inspire, resonate, and serve as a benchmark for leaders who are planning to bring scalable machine-learning solutions to their organizations.

The rising term of MLOps has been around for a while. Similar to DevOps, where practices are set to standardize the software development lifecycle, MLOps aims to efficiently and reliably deploy and maintain machine learning models. Since data scientists are traditionally trained to research and develop models, a gap needs to be filled to make models predict in a timely manner and scalable way.

Drizly is an e-commerce marketplace that specializes in alcohol delivery. Before it got acquired by Uber in 2021, Drizly had ridden along the crazy waves of rapid scaling, which resulted in dramatic shifts in data science needs. Throughout the process, Drizly has tried and abandoned a variety of tools such as the workflow scheduling tool Azkaban and AWS Redshift database, because of their inaccessibility (more on this later). It took us about two years of trial and failure…

Related posts:

As the emergence of digital technologies has transformed the news industry, the way that journalism is taught and learned in the college classroom and in campus-based journalism laboratories, or news…

If you live or work in Delhi, chances are you have ridden an elevator powered by gearless machines. These elevators, offered by companies such as Polo Elevators, are becoming increasingly popular in…

Now when many of us are spending more time at home (quarantined/working from home), many couples suddenly face new challenges in their relationships. All of a sudden, we don’t get the chance to “miss…