ML systems rarely scale as one unit. The model server, feature store, retrieval index, streaming pipeline, training job, cache, and monitoring stack all have different bottlenecks. Scaling only the visible bottleneck often moves the problem somewhere else.

A complete answer covers the whole system's scaling dimensions, not just the model server: QPS, candidate count, feature count, model cost, data volume, freshness, label delay, and traffic shape. Each dimension stresses a different component.