Data stream processing plays a critical role in providing fundamental statistics for various applications, such as anomaly detection. Still, the unbalanced distribution of data streams severely affects the performance of related algorithms, which motivates the recent studies on filter structure design to enhance the existing algorithms for a more precise estimation result. However, these filters are mainly designed for frequency-based filtration, while none of them can conduct universal filtration; apparently, frequency is not the only targeted metric in practical processing tasks, metrics like cardinality and persistence are of equal importance. To cope with the issue, we propose a novel filter framework to implement universal, lightweight, and accurate filtration. The filter framework is called Coupon Filter due to the interpretation of its flow-level filtration as a coupon collection process. We prove the filtration efficiency of our filter design and formally analyze its recording process. We deploy our filter on the three typical metrics (frequency, cardinality, and persistence) to illustrate its advantages. The experimental results on real Internet traces demonstrate the effectiveness of our filter in enhancing existing stream processing approaches in terms of accuracy and throughput. All source codes are available at Github https://github.com/duyang92/coupon-filter-paper.