Sampling algorithms in data stream environments

Data streams are large data sets generated continuously and at a fast tempo. Their arrival rate is large compared to the treatment and storage capacities. Thus, these streams cannot be entirely stored. That is why we need to treat them in a single pass, without storing them exhaustively. However, for a particular stream, it is not always possible to predict in advance all of the processing to be performed. It is therefore necessary to save some of this data for future treatments. These stored data then build “summaries”. Several ways exist for the construction of the summary, among them, the sampling algorithms. We propose in this paper an in-depth study of sampling methods used for the construction of data stream summaries. This paper includes two main parts. First, we introduce the basic concepts of data stream: Windowing models over data stream as well as data stream applications. Then we describe the different sampling algorithms used in stream environments. We particularly focus on their advantages and drawbacks. Finally, we compare the performance of the Simple Random Sampling to the chain sampling algorithm and we discuss the relevant research challenges for data stream sampling.