A model is only as good as the features it sees, and it must see the same features in training as in production. Feature pipelines compute values such as "orders in the last 30 days" from raw events. An offline store keeps their history for training; an online store keeps the latest value per entity for low-latency serving.
Two traps dominate. Training/serving skew happens when the two paths compute a feature differently, for example in different code or with different time windows. Leakage happens when a training example uses data from after the moment the prediction would have been made, which makes offline scores look great and production results disappointing. Point-in-time joins, which take each feature's value as of the example's timestamp, prevent leakage.
Compute features once for both paths, serve them fast and join them without leaking the future.