Nowadays, huge amounts of data are continuously created from different sources in different formats, and stored into data lakes for further use. A typical use case is to conduct online analyses in order to gain business insights and make better decisions. However, there are many obstacles to overcome when it comes on analysing Big Data online, such as dealing with schema-free data, conciliating different data formats, managing different locations, and allowing BI professionals and analysts to create their analytical data by themselves. In this paper, we propose some solutions to overcome these obstacles. We propose a data lake metadata model as well as a metadata-driven approach to create OLAP cubes from data lakes on-demand and in a self-service manner. We apply our work to Twitter social network and we present a proof-of-concept dedicated application.