Probabilistic automated language learning for configuration files

Abstract

Software failures resulting from configuration errors have become commonplace as modern software systems grow increasingly large and more complex. The lack of language constructs in configuration files, such as types and grammars, has directed the focus of a configuration file verification towards building post-failure error diagnosis tools. In addition, the existing tools are generally language specific, requiring the user to define at least a grammar for the language models and explicit rules to check. In this paper, we propose a framework which analyzes datasets of correct configuration files and derives rules for building a language model from the given dataset. The resulting language model can be used to verify new configuration files and detect errors in them. Our proposed framework is highly modular, does not rely on the system source code, and can be applied to any new configuration file type with minimal user input. Our tool, named ConfigC, relies on an abstract representation of language rules to allow for this modularity. ConfigC supports learning of various rules, such as orderings, value relations, type errors, or user defined rules by using a probabilistic type inference strategy and defining a small interface for the rule type.