Accrual Failure Detectors

Failure detection is a fundamental building block for ensuring fault tolerance in distributed systems. For this reason, many people have been advocating that failure detection should be provided as some form of service [1, 2, 3, 4, 5], similar to IP address lookup (DNS) or time synchronization (e.g., NTP). Unfortunately, in spite of important technical breakthroughs, this view has met little success so far. We believe that one of the main reasons is the conventional binary interaction (i.e., trust vs. suspect) that makes it difficult to meet the requirements of several distributed applications running simultaneously. For this reason, we advocate a different abstraction that helps decoupling application requirements from issues related to the underlying system. It is well-known that there exists an inherent tradeoff between (1) conservative failure detection (i.e., reducing the risk of wrongly suspecting a running process), and (2) aggressive failure detection (i.e., quickly detecting the occurrence of a real crash). Thre exists a continuum of valid choices between these two extremes, and what defines an appropriate choice is strongly related to application requirements.