In beam-based massive multiple-input multiple-output systems, signals are processed spatially in the radio-frequency (RF) front-end and thereby the number of RF chains can be reduced to save hardware cost, power consumptions and pilot overhead. Most existing work focuses on how to select, or design analog beams to achieve performance close to full digital systems. However, since beams are strongly correlated (directed) to certain users, the selection of beams and the scheduling of users should be jointly considered. In this paper, we formulate the joint user scheduling and beam selection problem based on the Lyapunov-drift optimization framework and give the optimal scheduling policy in a closed-form. In order to make the scheduling decisions based only upon statistical channel state information, asymptotic expressions of the downlink broadcast channel capacity are derived. To address the weighted sum rate maximization problem, an algorithm based on block coordinated update is proposed and proved to converge to the global optimum of the relaxed problem. To further reduce the complexity, an incremental greedy scheduling algorithm is also proposed, whose performance is proved to be bounded within a constant multiplicative factor of $(1-e^{-1})$. Simulation results based on widely-adopted spatial channel models are given, which show that the proposed scheme is close to optimal, and outperforms the state-of-the-art beam selection schemes.