Modern single-pass algorithms are becoming sufficiently sophisticated so that they can now, in some cases, approach the performance of multiple pass algorithms. “Perhaps the big surprise of streaming is how much is possible with only one look at the data,” says Cormode. “We can approximate the entropy, find how many distinct items there are, or identify the most frequently occurring items.” These tasks are relatively straightforward with a traditional multiple-pass approach, but accomplishing the same objectives with a single-pass architecture—while managing system costs—calls for more sophisticated techniques. “The cost of these algorithms depends on how accurately the answer is approximated,” Cormode says. “It turns out that we can get accurate answers with only kilobytes of storage.