Regular Programming over Data Streams

  • Mukund Raghotaman | University of Pennsylvania

The problem of programming stream transformations is becoming increasingly important with the emergence of large data streams such as from sensors, internet routers, and gene sequences. Many monitoring queries—for example, what is the average number of packets transmitted by a router per second—and policy actions depend on temporal properties, such as recognizing certain patterns of events, and iteratively aggregating results from these patterns. In this presentation, I will describe the model of quantitative regular expressions (QREs), a simple formalism to conveniently express such queries. Starting with a small set of base functions, QREs are built using function combinators such as choice, concatenation, and iteration, similar to how regular expressions are combined using union, concatenation and Kleene-*. The framework is parameterized by the cost operations of interest, such as max, min, sum, average and rank, and even generalizes to non-numerical output domains such as strings. Appropriate expressiveness results hold as a function of the chosen cost operations. QREs are an exciting framework because they have a well-understood theory, and can be easily compiled into fast streaming evaluation algorithms. In many cases, memory usage is bounded by a constant, and in some other cases, accurate queries can be automatically converted into approximate function evaluators. I will conclude by describing potential applications in monitoring cyber-physical systems and other data stream sources and future work into automatically parallelizing stream queries.

Series: Microsoft Research Talks