In this paper, we present a mechanism for scheduling multi-class traffic in a node of a packet-based communication network. We focus in particular on the output queues used for contention resolution at the output links of a network node. The mechanism, referred to as R-scheduling, is based on the use of in-queue reservations for future arriving data packets. Each traffic class has a given number of class-dedicated reservations at its disposal. If a data packet is enqueued, it seizes its reservation which is closest to the head of the queue by taking its place in the queue, after which a new reservation belonging to the same traffic class as the enqueued packet is created at the tail of the queue. The performance of the R-scheduling mechanism is investigated by means of extensive simulations for a wide parameter space. Based on these simulations, a meta-model is constructed that allows to translate given QoS (Quality of Service) targets into working parameters for the R-scheduling mechanism. We show that R-scheduling offers unique properties in terms of the QoS delivered to each of the traffic classes. In particular, it allows shaping the distributions of the queueing delay perceived by each traffic class in order to achieve a target spacing of their delay quantiles. Other important features that can be effectuated by R-scheduling are: isolation of traffic classes through temporal priority, mitigation of packet starvation for lower-priority traffic classes and early provision of tight delay bounds. The mechanism has low complexity and is suitable for dynamic applications.