Flow size estimation is a key task in per-flow measurement. In real scenarios, among millions of flows, people are often particularly interested in a small subset of given flows. These flows with high priorities are called important flows. In this poster, we propose the Cuckoo sketch, which provides higher accuracy for important flows. The key idea of the Cuckoo sketch is to separate the important flows from the unimportant ones and store them in different structures: the important flows are stored in a Cuckoo hash table where the exact size of the flow is stored while the unimportant flows are stored in a Count-Min sketch. The reallocation mechanism of the Cuckoo hashing is used to make some space in the Cuckoo hash table for the important flow by evicting the ones that are less important, which leads to a slight loss in throughput. Experimental results show that Cuckoo sketch provides better accuracy on important flows, reducing the average relative error by up to 69%, with only suffering from a negligible loss in accuracy for unimportant flows.