The hardest production problems in modern recommendation systems often sit outside the machine-learning model, according to an engineering presentation published by InfoQ. Former Twitter, Walmart and Netflix engineering leader Mallika Rao argues that retrieval, ranking, inference, evaluation and feedback must be treated as one adaptive distributed system.
Older recommenders could retrain models and refresh features on a schedule. Newer systems ingest behavior continuously and may change while a user is still interacting with a product. That makes data freshness and end-to-end latency part of model quality: a strong ranker cannot rescue stale candidates, and a fast retrieval stage does not help if later inference exhausts the response-time budget. Coordination failures at component boundaries can be more damaging than weaknesses inside an individual model.
Rao recommends treating evaluation as a first-class production capability rather than relying only on offline accuracy scores. Teams need to observe the complete feedback loop, test how changes affect real behavior and account for cost, compliance and user trust. The lesson also applies to AI agents, which increasingly learn from live context and operate under uncertainty. Before adding a more capable model, engineers should verify that the surrounding data, orchestration and measurements can reveal whether the system actually improved.