Concord: Learning Network Configuration Contracts

Misconfiguration is frequently cited as a leading cause of service disruptions and outages. To prevent misconfiguration, we introduce network contracts—lightweight configuration checks that run efficiently, localize errors to specific lines, and require no heavyweight modeling of network protocols. We develop a tool Concord to learn contracts automatically from example network configurations. By checking these learned contracts against new or changed configurations, Concord finds likely configuration bugs before they can impact the network. Key to our approach is a scalable algorithm for learning "relational" contracts that capture complex dependencies between configuration settings. We deployed Concord as part of a cloud-based configuration management service and evaluated its scalability, coverage, precision, and utility on two large real-world configuration datasets.