We cache recommendations to reduce server load. Could that actually be hurting our engagement numbers over time?
Yes, caching recommendations can degrade engagement over time if the system is not notified of repeated exposures. When a recommender receives no feedback on items it has already shown, it loses the signal that non-engagement carries - treating a skipped item identically to an unseen one. The result is a recommendation loop where items continue appearing despite repeated non-engagement. The fix is ensuring every cached impression is logged back to the recommender so it can factor repeated exposure into future ranking decisions.