A mobile-money fraud model uses "transfers in the last hour". Training computes it one way from old logs, the live app another way from the stream, and the two quietly disagree. Put one feature store between them, then let the training join peek at the future, and let the cache go stale. This lesson needs JavaScript.