Abstract

While the prevailing approach to the big data challenge is to scale up/out the computation, the other, complimentary approach is to scale down the data, which is highly effective when small errors in the results can be tolerated. Random sampling is one of the main tools to scale down the data, and has been well studied in both computer science and statistics. However, as modern big data systems continue to evolve, they put new constraints and requirements on the sampling methods. This talk will present some recent results on random sampling in large-scale database systems, including random sampling over data streams, in a distributed system, on spatial data, and sampling for SQL queries.