Confessions of an ML Infra Engineer

Cookpad is the world’s largest recipe sharing platform, welcoming 100 million monthly active users. Our platform was designed for global scale, but was not always designed with MLOps in mind.

We set up our Machine Learning Infrastructure team in 2018 to work alongside the established Machine Learning Research team to build machine learning driven features in the Cookpad app.  As ML Infrastructure engineers our ideal architecture had a long wish-list for MLOps (model versioning, data provenance, automatic retraining, an A/B testing framework, the list goes on...) But ivory tower architecture was not going to get us very far - we were joining an established platform!

Our journey towards MLOps has been incremental - we have collaborated with other engineering teams and product managers to run experiments and release machine learning services, all whilst navigating the rapidly evolving landscape of MLOps tools and technologies.      In this talk we will share some stories from our two years of ML infra at Cookpad - some success stories, some lessons learned and some insights into how far we have made it towards our goal of a gold-plated MLOps platform.