There are several replica control algorithms for managing replicated files in the face of network partitioning due to site or communication link failures. Pessimistic algorithms ensure consistency at the price of reduced availability; they permit at most one (distinguished) partition to process updates at any given time. The best known pessimistic algorithm, voting , is a “static” algorithm, meaning that all potential distinguished partitions can be listed in advance. We present a dynamic extension of voting called dynamic voting . This algorithm permits updates in a partition provided it contains more than half of the up-to-date copies of the replicated file. We also present an extension of dynamic voting called dynamic voting with linearly ordered copies (abbreviated as dynamic-linear ). These algorithms are dynamic because the order in which past distinguished partitions were created plays a role in the selection of the next distinguished partition. Our algorithms have all the virtues of ordinary voting, including its simplicity, and provide improved availability as well. We provide two stochastic models to support the latter claim. In the first (site) model, sites may fail but communication links are infallible; in the second (link) model the reverse is true. We prove that under the site model, dynamic-linear has greater availability than any static algorithm, including weighted voting, if there are four or more sites in the network. In the link model, we consider all biconnected five-site networks and a wide variety of failure and repair rates. In all cases considered, dynamic-linear had greater availability than any static algorithm.

In a voting-based algorithm, a replicated file can be updated m a partltlon if It contams a maJorlty of copies. In this paper, we propose an extension of this scheme which permits a file to be updated m a partition provided it contams a maJorlty of up-to-date copiesOur scheme not only preserves mutual consistency of the replicated file, but provides improvement m its avallablllty as wellWe develop a stochastic model which gives msrght into the rmprovements afforded by our scheme over the voting scheme I INTRODUCTIONA partttronzng of a distributed database (DDB) occurs when the sites m the network split into groups of commumcatmg sites due to node or commumcatlon farluresThe sites in each group can communicate with each other, but no site m one group 1s able to communicate with sites m other groups We refer to each such group as a partrtzoltThe algorithms which allow a partltloned DDB to continue functronmg generally fall into one of two classes + Those m the first class take a pesszmzsttc approach m that they share the philosophy that mutual consistency 1s of conslderably greater importance than avallablhty [l, 2,3,8,11,13,18,25,29] Consistency 1s enforced by t There are algorithms (see [E&15,23,28] for example) which do not belong to either of these two classes, however, they require a prlor~ knowledge of the kmd of updates to be made to the file We make no such assumption m this paper Permission to copy wlthout fee all or part of this material 1s granted provided that the copies are not made or dlstrlbuted for direct commercial advantage, the ACM copyright notice and the title of the pubhcatlon and Its date appear, and notlce 1s given that copymg IS by permission of the Assoclatlon for Computmg Machmery To copy otherwise, or to repubhsh, reqmres a fee and/or specfic permission permitting files to be accessed only m one partition As a consequence, any updates which are permltted m a partition do not conflmt with updates m other partltlons, assuring mutual consistency of data when partltlons are reunitedThe algorithms m the second class take the approach that the database must be available even m the face of node or commumcatlon link failures and permit every group of sites u-r a partltloned DDB to perform new updates [9,20,22,30,31] Since now the databases of the partitions will clearly diverge, they require a strategy for conflict detection and resolutton Usually, rollbacks are wed as a means for preserving consistency, conflmtmg transactions are rolled back when partitions are reunited Since coordmatmg the undoing of transactions IS a very difficult task, these methods are called optrmzstzc since they are useful prlmarlly in a sltuatlon where the number of items m a particular database IS large and the probabmty of conflicts among transactions is smallIn this paper, we consider only the pesslmlstlc strategies Although these methods guarantee that there will not be more than one partition performing the update at the same time, they cannot guarantee that there will be a partition which can perform updates ...

No abstract

There are several consistency control algorithms for managing replicated files in the face of network partitioning due to site or communication link failures. In this paper, we propose a hybrid scheme that integrates the static voting protocol and dynamic voting with linearly ordered copies.We use a stochastic model to compare the file availability afforded by the proposed hybrid scheme against the availabilities of voting, dynamic voting, and dynamic voting with linearly ordered copies. The analysis provides evidence for the conjecture that the hybrid scheme is the optimal algorithm in the context of the stochastic model.