Diagnosis of Service Failures by Trace Analysis with Partial Knowledge

The identification of the source of a fault (“diagnosis”) of orchestrated Web service process executions is a task of growing importance, in particular in automated service composition scenarios. If executions fail because activities of the process do not behave as intended, repair mechanisms are envisioned that will try re-executing some activities to recover from the failure. We present a diagnosis method for identifying incorrect activities in service process executions. Our method is novel both in that it does not require exact behavioral models for the activities and that its accuracy improves upon dependency-based methods. Observations obtained from partial executions and re-executions of a process are exploited. We formally characterize the diagnosis problem and develop a symbolic encoding that can be solved using constraint solvers. Our evaluation demonstrates that the framework yields superior accuracy to classic dependencybased debugging methods on realistically-sized examples.