An important goal of self-managing databases is the autonomic adaptation of the database configuration to evolving workloads. However, the diversity of SQL statements in real-world workloads typically causes the required analysis overhead to be prohibitive for a continuous workload analysis. The workload classification presented in this paper reduces the workload analysis overhead by grouping similar workload events into classes. Our approach employs clustering techniques based upon a general distance function for DBS workload events. To be applicable for a continuous workload analysis, our workload classification specifically addresses a stream-based, lightweight operation, a controllable loss of quality, and self-management.