A delay-tolerant network (DTN) is characterized by an intermittent topology where nodes and links may be down for long periods of time. This work evaluates whether it is more efficient to stop and resume bundle transmissions on the next contact opportunity after they become interrupted by an ending contact (preemptive model) than dropping the affected bundles (lossy model). Common intuition tells that the former model involves higher response times than the latter, but also fewer losses, so the optimal choice may not be immediately evident. For the analysis, stochastic models that describe the bundle transmission process over a DTN channel that is characterized by recurrent and random contact opportunities are formulated. Simulation results confirm the accuracy of the models. The study reveals the performance trade-offs between the two design choices and indicates that, contrary to expectations, the lossy model makes more efficient use of the channel than with preemption according to Kleinrock's power metric, which integrates throughput, delay, and loss. However, after considering retransmissions, preemption yields slightly higher power than the lossy approach. This work brings new insights into DTN performance and design choices.