Mining Web Logs to Debug Distant Connectivity Problems

Content providers base their business on their ability to receive and answer requests from clients distributed across the Internet. Since disruptions in the flow of these requests directly translate into lost revenue, there is tremendous incentive to diagnose why some requests fail and prod the responsible parties into corrective action. However, a content provider has only limited visibility into the state of the Internet outside its domain. Instead, it must mine failure diagnoses from available information sources to infer what is going wrong and who is responsible. Our ultimate goal is to help Internet content providers resolve reliability problems in the wide-area network that are affecting enduser perceived reliability. We describe two algorithms that represent our first steps towards enabling content providers to extract actionable debugging information from content provider logs, and we present the results of applying the algorithms to a week’s worth of logs from a large content provider, during which time it handled over 1 billion requests originating from over 10 thousand ASes.