We analyze the delay experienced in a discrete-time priority queue with a train-arrival process. An infinite user population is considered. Each user occasionally sends packets in the form of trains: a variable number of fixed-length packets is generated and these packets arrive to the queue at the rate of one packet per slot. This is an adequate arrival process model for network traffic. Previous studies assumed two traffic classes, with one class getting priority over the other. We extend these studies to cope with a general number M of traffic classes that can be partitioned in an arbitrary number N of priority classes (1 ≤ N ≤ M). The lengths of the trains are traffic-class-dependent and generally distributed. To cope with the resulting general model, an (M × )∞-sized Markovian state vector is introduced. By using probability generating functions, moments and tail probabilities of the steady-state packet delays of all traffic classes are calculated. Since this study can be useful in deciding how to partition traffic classes in priority classes, we demonstrate the impact of this partitioning for some specific cases.