News recommendation is an area of research where we deal with a non-stationary source of documents which are recommended to the users of the publishers' web sites. Predominant success metric is the attention span of a user expressed in terms of time spent on site and number page views. The key modeling problem is the fact that the most relevant news to be recommended are usually the fresh ones having no usage history, ie. the goal is to recommend items about which we don't know much. There are several types of data one considers when doing news recommendation. The most obvious ones are content of the articles and collaborative filtering with the help of contextual features like GeoIP, time, and demographics. More sophisticated types of data include semantics extracted from the text, meta data and inferred demographics (look-a-likes). Once having a representation determined, an important dimension is granularity of modeling for personalized information delivery balanced with the required response time (processing speed). In this contribution we will present a solution using most of the above ingredients built for a large online business news providers with up-to few hundred page views per second. The talk will focus on design decisions leading to a successful self adaptive system serving millions of users per day.