Abstract

Today's distributed systems need to replicate data for scalability, performance
and availability. However when there are write operations to the replicated data,
it is impossible to immediately and synchronously update all replicas, forcing
systems to chose between fast writes and data consistency. Additionally these
distributed systems also suffer from network partitions which can result in the
replicas on opposite sides of the partition
diverging. To deal with these problems and tradeoffs, different systems provide
different consistency models for these replicas - from strong models like
serializability to weak models like eventual consistency - which vary in the
guarantees they provide about the consistency of data read from different replicas.
The choice of consistency models is ultimately based on application requirements
and some distributed systems or frameworks for distributed systems would even go
to the extent of providing tuning options to provide different consistency models
within the same distributed system.