Mobile edge computing (MEC) has been a promising architecture for providing delay-sensitive computing services to mobile users. In the provisioning of MEC services, one big issue is the potential mismatching between dynamic task arrivals and resources allocated for different types of applications at edge servers, which can cause degraded system performance. In this paper, we study the joint optimization of application deployment and task assignment at different time scales in a MEC system constituent of multiple cloudlet service providers. The design objective is to maximize the system profit while meeting the task delay requirements of different applications. We formulate this problem as a mixed integer linear programming (MILP) problem. For the online scenario where future task arrival information is unknown in advance, we propose a multi-armed bandit based application deployment and maximum flow matching based service coalition algorithm. We deduce the complexity of the proposed algorithm and prove that it satisfies individual rationality and coalition rationality. Extensive simulations are carried out and the results show that the proposed algorithm can effectively improve the system profit while satisfying the task delay requirements.