讲座题目：Tradeoffsbetween consistency and performancein geo-replicated systems
Cheng Li is an associate researcher at theINESC-ID, Technical University of Lisbon. Prior to this, he was appointed as asenior member of technical staff at Oracle Labs (Swiss), working on improvingthe performance and energy-efficiency of large-scale database-processingsystems. He pursued his PhD degree at the Max Planck Institute for SoftwareSystems and Saarland University in
As response time has a strong negativeimpact on user activities and revenue of online services, geo-replication hasbeen widely adopted for improving performance. However, there exists afundamental tension between consistency and performance in geo-replicatedsystems, i.e., weak consistency makes systems fast but likely introducesundesirable behaviors, while strong consistency precludes anomalies but slowsdown systems. To address this tension, some proposals from academia andindustry allow operations to run at different consistency levels. To allowthese levels to co-exist, we first propose a novel consistency model, RedBlueconsistency, under which commutative and invariant-preserving operations coulduse eventual consistency and the remaining ones must use strong consistency.RedBlue improves performance of a wide range of web applications, but it alsointroduces unnecessary and costly coordination to some other applications. Inorder to further improve the performance of geo-replicated systems, second, wegeneralize the intuitions and principles behind RedBlue consistency, andintroduce a new generic model, PoR consistency, where consistency semantics arecaptured by the visibility restrictions between pairs of operations.