Many sources nowadays produce stream data continuously and ubiquitously, with examples of web logs, online users’ click-streams, network traffic and dynamic transactions records. This imposed a great challenge in certain online interactive applications where real-time decisions by the system are needed to be made on-the-fly, given that the stream data are vast and response time requirement is stringent. It has already been identified in the research community that traditional data-mining, such as decision tree algorithms require scanning through a finite and structured set of data for learning, which may not be suitable in highly interactive web applications where the data come in streams. Hoeffding Tree Algorithm (HTA) that is a core component of Very-FastDecision-Tree is one of the promising data stream mining techniques for real-time applications. However, an important parameter in HTA called tree-splitting-error needs to be controlled for meeting the real-time requirements and satisfying the classification performance. This paper sheds some light into the relations between HTA and real-time constraints.