Wait-free Replicated Data Types and Fair Reconciliation

2025-08-25

AI Summary

Wait-free Replicated Data Types (RDTs) in distributed systems suffer from repeated operation rollbacks and client starvation. Method: We propose the first generic framework that simultaneously guarantees wait-freedom, stable convergence, and fair progress. It models client operations as a directed acyclic graph (DAG) and introduces a novel coordination function enabling local view exchange and monotonic state merging among replicas, thereby ensuring all replicas share a strictly increasing prefix of stable operations. Contribution/Results: This is the first work to achieve eventual consistency, strong availability, and fairness under wait-freedom: every client operation is stably committed within a bounded number of steps, eliminating perpetual starvation. Experimental evaluation demonstrates significant reductions in rollback rates, alongside improved throughput and response-time fairness under asynchronous concurrent workloads.

Technology Category

• Game Theory and Economic Paradigms: Fair Division
• Search and Optimization: Distributed Search
• Multiagent Systems: Coordination and Collaboration

Application Category

• Responsible Web: Data and user privacy-enhancing technologies for the Web
• Security and Privacy: Data transparency and provenance
• Systems and Infrastructure for Web, Mobile and WoT: Data management and stream processing for Web, mobile and wireless applications

Abstract

Replication is a standard way to maintain availability of shared data in fault-prone distributed systems. To make sure that the data replicas are up-to-date, they need to synchronize, which typically means engaging the replicas in waiting for coherent responses from each other. The amount of waiting depends on the consistency and availability guarantees we impose on the system. The folklore CAP theory states that strong consistency (the set of replicas create an illusion of one correct server) and strong availability (the replicas' states are reachable despite network partitions) cannot be implemented in the same system. A popular way to deal with this impossibility is to relax consistency to be only eventual: the replicas eventually converge to the same state. In return, the replicas can be wait-free, i.e., the clients can get the data from the closest replica without waiting for other ones. Wait-free data replication faces two important challenges. First, the operations issued by the clients may be constantly revoked, i.e., their effects can be repeatedly recomputed due to asynchrony and concurrency. Second, even if some operations eventually stabilize in their effects, a particular client may still experience starvation if, from some point onward, each of its operations is later revoked. In this paper, we address these challenges through a general DAG-based framework for replicated data types, where replicas exchange their local views and merge them using a reconciliation function. Within this framework, we design reconciliation functions that implement a wait-free eventually consistent replicated state machine ensuring both stable convergence and fair progress. Specifically, every replica maintains a growing sequence of client operations, and we guarantee that: (1) all replicas share a common, monotonically growing stable prefix of operations, and (2) no client starves.

Problem

• Designing wait-free replicated data types for distributed systems
• Ensuring stable operation convergence without constant revocation
• Providing fair progress guarantees to prevent client starvation

Innovation

• DAG-based framework for replicated data types
• Reconciliation function ensuring stable convergence
• Wait-free eventually consistent state machine