2026-10-14T04:52:29Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32641938 2026-10-01T16:19:34Z portal_1172 item_type_8 month_year_10_2026

An architecture for scaling performance and availability in global consensus systems Michael Davis (360648) PUREID: 659009234 Consensus replicated state machines distributed systems paxos scalability Modern global consensus systems require both high availability and performance scalability to support planet-scale applications. However, existing consensus algorithms treat these properties as fundamental trade-offs: scaling fault tolerance, adding servers, or geographic distribution inevitably degrades either performance or availability. This thesis identifies that this tradeoff stems from a core bottleneck in traditional Paxos and Flexible Paxos-based algorithms: the requirement for all agents to participate in all aspects of every agreement, called here global participation.<br><br>This thesis presents a systematic approach to removing global participation bottlenecks in consensus through Proposer Acceptor Mappings (PAMs), which partition consensus agents across different portions of the ledger. First, to enable broad adoption of PAMs, this thesis presents PAM-based Paxos, an abstraction generalisation of existing Paxos algorithms that unifies Synod and Multi-Paxos constructions whilst enabling agent partitioning. Building on this foundation, three PAM designs are presented: the Geographic PAM, Horizontal PAM, and Conflict Minimising PAM. Each addresses specific scalability bottlenecks arising from global participation. To further assist broad adoption, the thesis introduces the Proactive Proposals algorithm, to coordinate multiple proposer groups without critical-path synchronisation within a broad set of Paxos leadership paradigms. It avoids critical-path synchronisation across groups by regulating proposal rates through a two-tier pipeline mechanism based on observed learner execution rates.<br><br>Finally, the thesis demonstrates the practicality of PAM-based consensus through Derek, a leaderless Synod-based consensus algorithm implementing Proactive Proposals. Empirical evaluation across diverse deployments shows that Derek (1) maintains near-constant performance when scaling fault tolerance, (2) enables horizontal scaling throughput when adding servers in multi-datacenter consensus, and (3) achieves superior geographic scaling whilst simultaneously improving availability compared to existing geographic scaling approaches. These results demonstrate that PAM-based partitioning can decouple the traditionally coupled dimensions of fault tolerance, performance, and availability in global consensus systems. 2026-10-01T16:19:34Z Text Thesis 10.17034/32641938.v1 https://figshare.com/articles/thesis/An_architecture_for_scaling_performance_and_availability_in_global_consensus_systems/32641938 All Rights Reserved