Reporting partitions authoritatively (using SDNs) for more robust distributed systems

This paper addresses network partitions, a core challenge for both distributed systems and networking. The networking community aims to reduce the frequency of network partitions, by engineering ever more reliable networks. The distributed systems community aims to reduce the severity of network partitions, by designing robust application logic. Unfortunately, this application logic is typically complex, because of uncertainty: unresponsive processes may be crashed, slow, temporarily partitioned, or permanently partitioned. This paper explores a co-design of applications and the network in which the network tells applications that processes (not hosts) are permanently partitioned with certainty, which in turn simplifies application logic. We embody this approach in Hawk, a failure and partition detector that provides formal guarantees. These guarantees can simplify distributed applications; reduce the resources they require to tolerate failures; and improve their availability, by delivering failure notifications to them promptly.