Consistency without concurrency control in large, dynamic systems

摘要

Replicas of a commutative replicated data type (CRDT) eventually converge without any complex concurrency control. We validate the design of a non-trivial CRDT, a replicated sequence, with performance measurements in the context of Wikipedia. Furthermore, we discuss how to eliminate a remaining scalability bottleneck: Whereas garbage collection previously required a system-wide consensus, here we propose a flexible two-tier architecture and a protocol for migrating between tiers. We also discuss how the CRDT concept can be generalised, and its limitations.