Disclosed is a variable-size sampling method under a data-streaming environment, including: calculating a maximum window size that satisfies a lower limitation of a predetermined uniformity confidence level at all times; inputting a data stream to be sampled; comparing a data stream length input until a current time point with the maximum window size; inspecting a sample size and a sampling fraction if the maximum window size is larger than the data stream length; performing sampling by generating a slot to increase the sample size if the current sample size is smaller than a predetermined percentage (P %) of the data stream; and directly performing sampling without generating a slot if the current sample size is equal to or larger than the predetermined percentage (P %) of the data stream. As a result, degradation of uniformity confidence during variable-size sampling under a real-time streaming environment can be prevented to improve sampling performance.