Using volatile witnesses to extend the applicability of available copy protocols

Abstract

This paper presents a new replication control protocol tailored to environments where network partitions can only occur at a few well-defined partition points and are always the results of a gateway failure. The protocol implements the same 'write all read-one' rule as the available copy (AC) protocol. Unlike the AC protocol, this protocol divides servers holding copies into local servers that can communicate directly with each other and non-local servers that communicate with other servers through one or more gateways. While replicas stored on local servers are assumed to remain up to date as long as their server remains operational, replicas stored on non-local servers are required to maintain one or more volatile witnesses on the same LAN segment as the local servers and need to interrogate one of these witnesses before answering any user request.<>