The ever-growing amount of data available on the Internet can only be handled with appropriate personalization. One of the most popular ways to filter content matching users' interests is collaborative filtering (CF). Yet, CF systems are notoriously resource greedy. Their classical implementation schemes require a substantial increase in the size of the data centers hosting the underlying computations when the number of users and the volume of information to filter increase. This paper explores a novel scheme and presents DeRec, an online cost-effective scalable architecture for CF personalization. In short, DeRec democratizes the recommendation process by enabling content-providers to offer personalized services to their users at a minimal investment cost. DeRec achieves this by combining the manageability of centralized solutions with the scalability of decentralization. DeRec relies on a hybrid architecture consisting of a lightweight back-end manager capable of offloading CPU-intensive recommendation tasks to front-end user browsers. Our extensive evaluation of DeRec on reals workloads conveys its ability to drastically lower the operation costs of a recommendation system while preserving the quality of personalization compared to a classical approach. On average, over all our experiments, DeRec reduces the load on the back-end server by 72% when compared to a centralized alternative.