Replicated Data Management and Consistency Models
Stub. Reliability vs. availability; passive vs. active replication and fault-tolerant state machines; data-centric consistency models (strict, linearizability, sequential, causal) and client-centric models (eventual consistency, mobile clients); quorum-based implementation and replica placement; distributed shared memory; and case studies (Coda, Bayou, gossip-based replication).
Replication trades consistency, availability, and latency against each other, and this note maps the whole space: passive (primary-backup) vs. active (state-machine) replication as architectures, then the hierarchy of data-centric consistency models from strictest (strict consistency, linearizability) to progressively weaker (sequential, causal), and the client-centric models (eventual consistency, and the variants needed for mobile/disconnected clients) that trade consistency for availability under partition. Quorum protocols are the standard way to implement a chosen consistency level, and distributed shared memory is what happens when this consistency-model machinery is applied to give processes the illusion of a shared address space. Coda, Bayou, and gossip-based systems are worked case studies spanning this whole spectrum.
Outline (TODO — flesh out each)
- Reliability vs. availability
- Passive vs. active replication; fault-tolerant state machines
- Data-centric consistency models — strict, linearizability, sequential, causal
- Client-centric consistency — eventual consistency; consistency for mobile clients
- Implementing data-centric models: quorum-based protocols
- Replica placement
- Distributed shared memory
- Case studies: Coda, Bayou, gossip architecture