Abstract

Many current online services have stringent availability and performance requirements. High availabil-ity entails tolerating component failures and is typically accomplished with replication. For many online services, “high performance ” essentially means the ability to serve an arbitrarily large load, something that can be achieved if the service can scale throughput when provided with additional resources (e.g., processors). In light of the requirements of modern online services, combining high availability and high performance without sacrificing consistency is a challenging endeavor of utmost importance. Replication has been extensively studied by the research community. Early work in the database com-munity dates back to the late 1970s and early 1980s, addressing both theoretical and practical concerns (e.g., one-copy serializability [1], primary copy replication [10]). In the distributed systems community, foundational work on object replication (i.e., non-transactional behavior) dates back to the late 1970s (e.g., state machine replication [7]). Although the topic is well-established, general-purpose replication techniques are mostly used for high availability, not performance—in fact, a replicated service is usu-ally outperformed by its single-server implementation. Replication for performance typically sacrifices consistency: the behavior of the replicated service does not correspond to the intuitive behavior of a single-server service. Although some services can settle for weak guarantees (e.g., [4]), weak consistency