Modern distributed databases employ leader-based consensus protocols to achieve consistency, entailing certain trade-offs: typically either a scalability bottleneck or weak isolation. Leaderless protocols have been proposed to address these and other shortcomings of leader-based techniques, but these have not yet materialized into production systems.
This paper outlines compromises entailed by existing leaderless protocols versus leader-based protocols and proposes general techniques for addressing them. A new protocol, called Accord, is proposed with optimal failure tolerance that, under reasonable assumptions, achieves optimal latency of two message delays for transactions initiated by coordinators in any region, under any level of competing transactions and maximal tolerated process failures.
Benedict Elliott Smith will present the Accord protocol in our DistSys Reading Group. Benedict is an Apache Cassandra contributor with an interest in performance, correctness, and algorithm design.
When: February 9th 2022 at 2 PM EST (Check your time zone)