MLOps Community - System Design for RecSys & Search

Eugene Yan, who's built recommendation systems at places like Amazon and Alibaba, took the mic at an MLOps Community meetup to talk through how he actually approaches system design for RecSys and search. Not the usual conference fluff about embeddings and transformers in the abstract, but the plumbing: how these systems get architected, deployed, and kept running once real traffic hits them.

The talk and slides are now posted publicly, which matters more than it might seem. RecSys and search sit in an awkward spot in ML content. There's no shortage of papers on ranking models or retrieval algorithms, but far less on the boring, load-bearing decisions — how to structure candidate generation versus ranking stages, where caching belongs, how to think about latency budgets when a user expects results in under 200 milliseconds. Yan has built a reputation writing about exactly that gap on his own site, and this talk is an extension of that work into a live, Q&A-friendly format.

What's notable is the venue. MLOps Community isn't a research crowd; it's practitioners wrestling with deployment, monitoring, and the operational headaches of running ML in production. A talk on RecSys system design fits that audience well, since recommendation and search are two of the areas where the gap between a Jupyter notebook and a production system is widest — you're not just serving predictions, you're serving them at scale, under strict latency constraints, with feedback loops that can quietly wreck your model's performance over time.

There's no dramatic reveal here, no new framework being launched. It's a knowledge-sharing artifact: a recording, a slide deck, and a citation block for anyone who wants to reference it formally. But for engineers actually building these systems, that's often more valuable than another paper. Practical, battle-tested design patterns from someone who's shipped this stuff are rarer than they should be.