Economy and research increasingly depend on the timely analysis of large datasets to guide decision making. Complex analysis often involve a rich variety of data types and special purpose processing models. We believe, the database system of the future will use compilation techniques to translate specialized and abstract high level programming models into scalable low level operations on efficient physical data formats. We currently envision optimized relational and linear algebra languages, a flexible data flow language(A language inspired by the programming models of popular data flow engines like Apache Spark (spark.apache.org) or Apache Flink (flink.apache.org).) and scaleable physical operators and formats for relational and array data types. In this article, we propose a database system architecture that is designed around these ideas and we introduce our prototypical implementation of that architecture.