A Hierarchical Classification for Software Health Indicators

Experience shows that external failures of software systems are often preceded by deterioration in their internal state (i.e. an error). An error is defined as the difference between a computed, observed, or measured value or condition and the true, specified, or theoretically correct value or condition [1]. For software systems that are designed to degrade gracefully, the capability to provide a statistical indication of their internal well-being or health would be very valuable. For example, providing operators with advance warning allows them to take pre-emptive action before a major operational disruption occurs. Software health monitoring is an approach that strives for the early detection of internal errors in operational software systems [2]. This approach employs software health indicators to monitor particular facets of a target program’s execution. Each indicator derives its base monitoring information from one or more software sensors that collect data from specific parts of the internal program state. This paper proposes a hierarchical classification for software health indicators. The classification organises indicators into categories based on the number of information sources examined, as well as on how those sources relate to each other. The classification can be used to identify appropriate indicator classes for detecting errors and can serve as a guide for retrofitting health indicators for monitoring existing software. During the course of this research, the authors performed an extensive literature review of software monitoring systems and error detection techniques such as the use of assertions. The authors also gained hands-on experience deploying health indicators into several different target systems. This work resulted in the evolution of patterns for health indicators with common functionality and produced this classification. To date, more than 30 indicator classes have been identified under five main categories. The classification has provided a better understanding of the structure and diversity of indicators for health monitoring and helped the authors identify deficiencies in monitoring coverage. The classification can accommodate new indicator classes by extending existing categories or adding new ones.