Real-time data services can significantly increase, e.g., the profit of online trades by processing transactions within their deadlines. However, supporting the transaction timeliness in dynamic environments such as the WWW is challenging, since transactions may arrive in a bursty manner and execute longer than expected. As a result, the degree of data/resource contention may vary, causing deadline misses. To address this problem, we refine real-time data service performance metrics and develop a novel feedback-based scheme to manage the clean CPU utilization, which is the difference between the aggregate utilization and wasted utilization due to data conflicts and deadline misses, in main memory real-time databases. According to the control signal computed in the feedback loop, the QoS of relatively large transactions, which usually incur more data/resource contention, can be degraded under high contention. Admission control can also be applied to incoming transactions, if necessary, to improve the real-time database performance under severe overload. By carefully managing the clean utilization, we can substantially improve the success ratio, i.e., the fraction of the submitted transactions that have been admitted and finished within their deadlines. In a simulation study, which models bursty arrivals of long-running transactions with timing constraints, our approach improves the success ratio by up to an order of magnitude compared to existing approaches.