Mobile edge computing (MEC) has been regarded as a promising approach to deal with explosive computation requirements by enabling cloud computing capabilities at the edge of networks. Existing models of MEC impose some strong assumptions on the known processing cycles and unintermittent communications. However, practical MEC systems are constrained by various uncertainties and intermittent communications, rendering these assumptions impractical. In view of this, we investigate how to schedule task offloading in MEC systems with uncertainties. First, we derive a closed-form expression of the average offloading success probability in a device-to-device (D2D) assisted MEC system with uncertain computation processing cycles and intermittent communications. Then, we formulate a task offloading maximization problem (TOMP), and prove that the problem is NP-hard. For problem solving, if the problem instance exhibits a symmetric structure, we propose a task scheduling algorithm based on dynamic programming (TSDP). By solving this problem instance, we derive a bound to benchmark sub-optimal algorithm. For general scenarios, by reformulating the problem, we propose a repeated matching algorithm (RMA). Finally, in performance evaluations, we validate the accuracy of the closed-form expression of the average offloading success probability by Monte Carlo simulations, as well as the effectiveness of the proposed algorithms.