Datalog has been specifically designed for processing recursive queries. As such it seems the perfect ift for today's data analytics applications, which typically require some kind of iteration or recursion. However, basic analytical tasks such as computing shortest paths, betweenness centrality, or connected components in a graph as well as solving optimization problems by gradient descent or other methods crucially depend on aggregation. For Datalog, this poses new challenges due to the loss of the usual monotonicity properties in case of aggregation. In this talk, I will present recent work in which • we have studied a generalization of Datalog that allows for recursive computations over general semirings (with classical Datalog corresponding to the special case of the Boolean semiring), • we have analyzed the convergence of this generalization of Datalog, and • we have introduced a powerful new optimization technique that covers known optimizations such as magic-set rewriting as well as new ones.