Provenance and data differencing for workflow reproducibility analysis

摘要

One of the foundations of science is that researchers must publish the methodology used to achieve their results so that others can attempt to reproduce them. This has the added benefit of allowing methods to be adopted and adapted for other purposes. In the field of e-Science, services – often choreographed through workflow, process data to generate results. The reproduction of results is often not straightforward as the computational objects may not be made available or may have been updated since the results were generated. For example, services are often updated to fix bugs or improve algorithms. This paper addresses these problems in three ways. Firstly, it introduces a new framework to clarify the range of meanings of ‘reproducibility’. Secondly, it describes a new algorithm, PDIFF, that uses a comparison of workflow provenance traces to determine whether an experiment has been reproduced; the main innovation is that if this is not the case then the specific point(s) of divergence are identified through graph analysis, assisting any researcher wishing to understand those differences. One key feature is support for user-defined, semantic data comparison operators. Finally, the paper describes an implementation of PDIFF that leverages the power of the e-Science Central platform that enacts workflows in the cloud. As well as automatically generating a provenance trace for consumption by PDIFF, the platform supports the storage and reuse of old versions of workflows, data and services; the paper shows how this can be powerfully exploited to achieve reproduction and reuse. Copyright © 2013 John Wiley & Sons, Ltd.