Rapid growth in the cloud services market provides tremendous opportunities to cloud providers who have invested heavily in computing capacities but also has led, at time, to low utilization of capacities. To alleviate this problem, some providers have launched a low-priority service with preemptible (spot) instances, which allows them to attract more customers while keeping the right to reclaim capacities when necessary. In this study, we consider a provider who faces a heterogeneous pool of customers with fault-tolerant (interruptible) computing jobs. We develop an analytical framework that consists of a customer-choice model and a diffusion model to capture the underlying supply-demand dynamics and the resulting preemption probability. First, we examine a commonly used discount scheme for preemptible instances, namely, the uniform discount scheme, and derive the optimal discounted price, given customers' expectation of the preemption probability. Then, we propose another practical discount scheme, namely, the interruption-based discount scheme, which provides customers with compensation for interruptions. As long as the provider interrupts the preemptible instances randomly and customers are risk neutral, the two discount schemes are equivalent from the provider's perspective. That said, the proposed scheme is fairer than the uniform discount scheme from the customers' perspective, as the former provides more discounts to customers who experience more interruptions, whereas the latter provides all customers with the same discount. Finally, in the presence of risk-averse customers, through a numerical study, we find that the provider would be better off by adopting the uniform discount scheme in an environment in which the level of surplus capacity stays high and stable. Overall, however, the provider would be better off by adopting the proposed scheme when the level of the surplus capacity is moderate and volatile; the relative advantage of the proposed scheme enlarges as the average surplus capacity decreases and its volatility increases
