Correlated Independence: Why Redundant Storage Systems Share the Same Fate

Abstract

To maintain data availability and fault tolerance, redundant storage systems such as RAID arrays and distributed storage architectures are generally implemented. The usual approach assumes that redundancy by itself ensures independence, i.e., the failure of a single component does not coincide with the failure of other components. However, studies based on experience reveal that this assumption fails most of the time: components which are considered independent regularly show correlated failures, i.e., several components fail simultaneously due to a common cause or a combination of factors. Such factors may be common hardware designs, software bugs, firmware interactions, human operation mistakes, or environmental conditions such as temperature or power variations. These situations of correlation go against the functioning of redundancy and therefore risk the continuity of data centers, cloud infrastructures, and other crucial data systems where the service continuity and data integrity are mandatory. For this purpose, analysis of failure events was performed statistically, various storage installations were evaluated for correlation, and a detailed study of the case was done to identify the examples of synchronized failure from the present. The study states that the correlation between “independent” storage components is higher than the conventional assumption, thus fault-tolerance models are challenged. Here, the emphasis is laid on the necessity of revising the concept of redundancy, incorporating correlated failure models into system designs and risk assessments, and upgrading the predictive tools for the reliability of storage. Therefore, taking into account correlated failures is a prerequisite for the development of storage systems that are really self-supporting, so that redundancy not only guarantees protection but also avoids inducing a false feeling of ​‍​‍​‍‌security.