Highly-available data services for UNIX client-server networks: Why fault-tolerant hardware isn't the answer

Abstract

High-Availability, or HA, often tops the feature “wish-list” for customers putting mission-critical applications on-line. The meaning of HA is often imprecise, however. There is a common perception that HA is actually secondbest to “fault-tolerance” — identified with hardware redundancy, and perceived to depend on complex, proprietary, costly technology. The perceived requirement for a “fault-tolerant” machine arises from an erroneous focus on single machine availability and on hardware faults as the dominant issues for service availability. On the other hand, empirical studies reveal that software faults and planned administrative procedures are the dominant issues; and that a customer's real HA requirement comes down to a need for HA data access from a client network; i.e. service that is available despite software faults, hardware faults, scheduled maintenance, software upgrade, etc. Recent UNIX-based implementations of an HA configuration, based on dual-hosted disks, have demonstrated that HA data service is achievable using commodity UNIX hardware and software components. We illustrate with a description of a Highly Available Data Facility prototype we implemented in 1992. We compare our approach to other contemporary approaches to HA client-server computing.