Fundamentals of distributed system observation

Abstract

It's difficult to determine event order in distributed systems because of the observability problem. The author discusses this problem and evaluates different strategies for determining arrival order. The author analyzed four time stamping methods to determine their effectiveness in contending with observability problems. Although he focuses on distributed systems, the concepts also apply to any system exhibiting concurrency-the appearance of two or more events occurring simultaneously-including multiprocessor machines and uniprocessor multitasking. Events in this context may be the execution of single machine instructions or entire procedures; the level of granularity is unimportant. To define event order, the author uses the idea of causality-the ability of one event to affect another-because it allows us to reason independent of any particular time frame.