Data consistency

Data consistency is the property of a distributed or replicated system that ensures all participants observe the same sequence of state changes, or — in weaker formulations — that their observations converge to a common state over time. It is not a binary property but a spectrum, ranging from linearizability (every operation appears to execute instantaneously at a single point in time) through sequential consistency (operations appear to execute in some global order consistent with each process's own order) to eventual consistency (if no new updates are made, all replicas eventually converge).

The consistency model chosen for a system determines what guarantees programmers can rely on. Strong consistency models simplify reasoning — a program behaves as if there is only one copy of the data — but they require coordination, typically through consensus protocols or quorum systems, which imposes latency and availability costs. Weak consistency models sacrifice immediate agreement for performance and partition tolerance, pushing the burden of conflict resolution onto application logic.

The taxonomy of consistency models is large and often confusing. Causal consistency guarantees that causally related operations are observed in order, while concurrent operations may be observed in different orders by different processes. Session consistency guarantees that a single client sees its own writes in order. Monotonic reads guarantees that a process never observes an older value after having observed a newer one. These models are not arbitrary; they form a lattice ordered by strength, with linearizability at the top and the empty guarantee at the bottom.

The deeper significance of consistency models is that they encode theories of causality. A system with strong consistency assumes that there is a single correct order of events. A system with causal consistency assumes only that causes precede effects. The choice between them is a choice about what kind of universe the system operates in — one with a global clock, or one with only local causality.

The database industry's obsession with strong consistency is a historical accident inherited from single-node systems, where consistency was free. In distributed systems, consistency is never free; it is purchased with latency, availability, and complexity. The systems that will dominate the next decade are those that abandon the fantasy of a single global truth and instead embrace the reality of local, causal, eventually convergent narratives. Strong consistency is a luxury good, and the future belongs to systems that can do without it.