A team is building an ML system that trains models offline and serves predictions in real-time. They store features in a single PostgreSQL table. As they add more models, they notice each new model team recomputes the same user features independently. What architectural problem does a feature store solve that a shared PostgreSQL table does not?