OSCAR: a system for weak-consistency replication

Abstract

Database replication is an important technique for constructing viable distributed systems. It can be used to improve system performance, availability, and survivability. In recent years, the proliferation of intemetworks and distributed applications has increased the demand for geographically-distributed replicated databases. Typical database replication systems, which were designed for local area networks, are not well suited for these applications. Intemetworks introduce new problems because they often contain high-latency , low-bandwidth links and complicated fault modes including transmission failures, node failures, and network partitions.A replication system contains mechanisms for update propagation and interreplica consistency. The update propagation mechanisms must ensure that updates are efficiently and reliably propagated to all of the replicas. Reliable propagation is more difficult to achieve in an internetwork because of the limited capabilities of the links, the increased frequency of failures, and the new kinds of failures that can occur, especially network partitions. The consistency mechanisms must maintain consistency among the database replicas and balance application requirements for database consistency and availability.The traditional approach of strong consistency, which requires all available copies of a data item to have the same value, is difficult to achieve across an internetwork. Executing atomic commitment protocols over slow, unreliable links can be very inefficient, causing transactions to suffer long delays and reduced throughput.For example, a two-phase commit can incur a delay of more than a second over a satellite link. Common techniques for maintaining strong consistency, such as primary copy [ 11 and quorum consensus [ 2 J , severely limit availability during a network partition because they allow a particular data item to be updated in at most California one partition and they allow reads of only the most recent values. Furthermore, the performance of strong-consistency replication techniques does not scale well to large numbers of replicas.In contrast to strong consistency, a database replication system that provides weak consistency permits greater availability by allowing temporary inconsistencies to develop among the replicas. These temporary inconsistencies are the result of multiple users independently updating different copies of a database, as well as the time necessary for an update to propagate to all replicas. The system guarantees to resolve these inconsistencies and return the replicas to mutual consistency. Techniques for resolving inconsistencies can be based on the syntactic dependencies among conflicting transactions, the global order in which transactions were executed, or the semantic properties of database operations and applications. An example of the use of semantic information is the knowledge that certain data items can only be incremented or decremented. Weak consistency has been successfully used in an internetwork for airline reservation systems,...

Property-Based Testing: Climbing the Stairway to Verification

“…The refinement statement as shown in Figure 4 is largely boilerplate code. To generate this code, we designed a small domain-specific language (DSL), whose prototype has been implemented [27]. In this DSL, programmers can specify the function names, the definition of the abstraction function for the inputs and the refinement relation between the outputs, and other properties about the refinement statement, such as the determinism of the abstract function and whether the concrete function needs to operate under the IO monad.…”

Computer and Information Sciences

“…OSCAR (Open System for Consistency and Replication) [Downing90a,Downing90b] implements weak-consistency replication using a mixture of distributed and centralized elements. It provides reliable eventual message delivery, with a variety of message orderings.…”

The Adr Replication Manager

“…The basic idea is that updates will eventually reach all the replicas of a logical data item, similar to an infectious disease among the individuals of a population. 8 ' 10 The overhead of typical epidemic algorithms grows linearly with the number of physical data items 35 propose an improved epidemic algorithm which bounds the overhead to grow linearly with the number of data items that actually must be copied during propagation. Epidemic algorithms are useful for very large and heterogeneous distributed systems such as the Internet where the replicas of a logical data item may be spread over hundreds or even thousands of sites so that replicas will eventually (but not rapidly) converge towards a mutual consistent state.…”