A team serves real-time ML predictions. Their model requires 30 features. For each prediction request, the serving code runs 30 separate database queries (one per feature). p99 latency is 600ms. A colleague says "use a feature store." What specific problem does the online store component solve?