In overloaded task systems, it is by definition not possible to complete all tasks by their deadlines. However, it may still be desirable to maximize the number of in-time task completions. The performance of on-line schedulers with respect to this metric is investigated here. It is shown that in general, an on-line algorithm may perform arbitrarily poorly as compared to clairvoyant (off-line) schedulers. This result holds for general task workloads where there are no constraints on task characteristics. For a variety of constrained workloads that are representative of many practical applications, however, on-line schedulers that do provide a guaranteed level of performance are presented.