We consider communication-efficient weighted and unweighted (uniform) random sampling from distributed streams presented as a sequence of mini-batches of items. We present and analyze a fully distributed algorithm for both problems. An experimental evaluation of weighted sampling on up to 256 nodes shows good speedups, while theoretical analysis promises good scaling to much larger machines.
Communication-Efficient (Weighted) Reservoir Sampling
Lorenz Hübschle-Schneider and Peter Sanders