Storage Failures are not Independent. What can we do about it?

Many different errors and failures in storage have been studied quite extensively. A perhaps under-appreciated body of work has shown via statistical studies that successive failures and/or errors are demonstrably not independent. This has serious implications for how we model data loss in storage systems, because the conventional assumption is one of independent successive failures. The reason the independence assumption is still used is because we have few real ways of handling dependent events. This talk discusses the mathematical foundations of probabilistic dependence we have developed over the last two years and its applications to the storage industry and data protection models.

Learning Objectives:
1. Explore the literature documenting dependent failures and LSEs in storage
2. Learn about classes of probabilistic dependency structures
3. Discuss the implications for a stronger form of autocorrelation
4. Examine the impacts to data protection