Building a scalable Machine Learning feature store for DoorDash

They way they describe their store it seems like their inferences aren’t actually personalized except in aggregate.

Is there a 3rd source of user-specific features that are only sometimes used because DD knows what places I order from often and suggests them but can’t seem to remember that I always order roti and not naan when making suggestions. Is it a performance thing where user-specific lookups are costly?