期刊名称：Bulletin of the Technical Committee on Data Engineering
出版年度：2015
卷号：38
期号：1
出版社：IEEE Computer Society
摘要：It is well known that the ability to timestamp events, to keep track of the order in which they occur,or to make sure they are processed in a way that is meaningful to the application, is of paramountimportance in a distributed system. The seminal work by Leslie Lamport[30] has laid the foundationsto understand the trade-offs associated with the use of physically and logically synchronized clocksfor establishing the order of events. Lamport’s original work considered scalar clocks and has laterbeen extended to consider richer forms of logical clocks such as vector clocks and matrix clocks. Thepurpose of this paper is to revisit these concepts in the concrete setting of developing distributed datastore systems for the cloud. We look at how different clock mechanisms have been applied in practice tobuild different cloud data stores, with different consistency/performance trade-offs. Our goal is to gaina better understating of what are the concrete consequences of using different techniques to capture thepassage of time, and timestamp events, and how these choices impact not only the performance of clouddata stores but also the consistency guarantees that these systems are able to provide. In this processwe make a number of observations that may offer opportunities to enhance the next generations of clouddata stores.
