Date of Original Version
Copyright 2011 USENIX
Abstract or Description
Incremental processing of large-scale data is an increasingly important problem, given that many processing jobs run repeatedly with similar inputs, and that the de facto standard programming model (MapReduce) was not designed to efficiently process small updates. As a result, new systems specifically targeting this problem (e.g., Google Percolator, or Yahoo! CBP) have been proposed. Unfortunately, these approaches require the adoption of a new programming model, breaking compatibility with existing programs, and increasing the burden on the programmer, who now is required to devise an incremental update mechanism. We claim that automatic incremental processing of large-scale data is possible by leveraging previous results from the algorithms and programming languages communities. As an example, we describe how MapReduce can be improved to ef- ficiently handle small input changes by automatically incrementalizing existing MapReduce computations, without breaking backward compatibility or demanding programmers to adopt a new programming approach.
Proceedings of the USENIX Workshop on Hot Topics in Cloud Computing (HotCloud), 2011.