Résumé: Applications in domains ranging from large-scale simulations in astrophysics and biology to enterprise analytics rely on computational pipelines. A pipeline consists of modules and their associated parameters, data inputs, and outputs, which are orchestrated to produce a set of results. If some modules derive unexpected outputs, the pipeline can crash or lead to incorrect results. Debugging these pipelines is difficult since there are many potential sources of errors including: bugs in the code, input data, software updates, and improper parameter settings. We present BugDoc, a system that automatically infers the root causes and derive succinct explanations of failures for black-box pipelines. BugDoc does so by using provenance from previous runs of a given pipeline to derive hypotheses for the errors, and then iteratively runs new pipeline configurations to test these hypotheses. Besides identifying issues associated with computational modules in a pipeline, we also propose methods for: "opportunistic group testing" to identify portions of data inputs that might be responsible for failed executions (what we call ), helping users narrow down the cause of failure; and "selective instrumentation" to determine nodes in pipelines that should be instrumented to improve efficiency and reduce the number of iterations to test. Through a case study of deployed workflows at a software company and an experimental evaluation using synthetic pipelines, we assess the effectiveness of BugDoc and show that it requires fewer iterations to derive root causes and/or achieves higher quality results than previous approaches.

Langue: Anglais