Beaver: Practical Partial Snapshots for Distributed Cloud Services

Distributed snapshots are a classic class of protocols used for capturing a causally consistent view of states across machines. Although effective, existing protocols presume an isolated universe of processes to snapshot and require instrumentation and coordination of all. This assumption does not match today’s cloud services—it is not always practical to instrument all involved processes nor realistic to assume zero interaction of the machines of interest with the external world.

To bridge this gap, this paper presents Beaver, the first practical partial snapshot protocol that ensures causal consistency under external traffic interference. Beaver presents a unique design point that tightly couples its protocol with the regularities of the underlying data center environment. By exploiting the placement of software load balancers in public clouds and their associated communication pattern, Beaver not only requires minimal changes to today’s data center operations but also eliminates any form of blocking to existing communication, thus incurring near-zero overhead to user traffic. We demonstrate the Beaver’s effectiveness through extensive testbed experiments and novel use cases.