RecSys 2022 Keynote - Is the Juice Worth the Squeeze?

Eugene Yan took the stage at RecSys 2022's ORSUM workshop with a title that doubles as the whole argument: is the juice worth the squeeze? Online recommender systems promise real-time adaptation, but they drag in extra infrastructure, extra compute cost, and extra operational headache. Yan's talk was essentially a cost-benefit audit of that tradeoff, built from real work at Amazon Books rather than theory.

His core claim cuts against the usual hype. Online systems aren't just about responsiveness or plugging in fresh context, they can, somewhat counterintuitively, save money too. That's a different pitch than the standard "real-time is better because it's real-time" line, and it's the part likely to stick with engineers deciding whether to greenlight a streaming pipeline versus a nightly batch job.

To make the case concrete, Yan walked through three systems built at Amazon Books. One handles cold-start users, the notoriously hard problem of recommending well before you've collected any real signal on someone. Another folds in context, adjusting recommendations based on what's happening right now rather than relying on stale historical patterns. A third targets waste and staleness directly, trimming the inefficiencies that batch systems tend to accumulate when they only refresh on a schedule.

The honest bit, and the reason this talk is worth more than a highlight reel, is that Yan frames batch as the right default in plenty of cases. Online recsys aren't a universal upgrade. They're a tool you reach for when responsiveness, context, or cost genuinely tip in their favor, and the talk's whole structure is built around helping teams recognize which situation they're actually in before signing up for the added complexity.

Yan credits a long list of Amazon Books teammates for getting the work through approvals and into shareable shape, a small reminder that keynotes like this are usually the visible tip of a much larger internal engineering effort.