We propose a novel Drone-as-a-Service (DaaS) composition framework considering the recharging constraints and the stochastic arrival of drone services. We develop a service model and a quality model for drone delivery services. A skyline approach is proposed that selects the optimal set of candidate drone services to reduce the search space. We propose a heuristic-based multi-armed bandit approach to compose drone services minimizing delivery time and cost. Experimental results prove the efficiency of the proposed approach.