Eris: Coordination-Free Consistent Transactions Using In-Network Concurrency Control
(link) Jialin Li, Ellis Michael, Dan R. K. Ports (University of Washington)
- What is the problem? Strong consistency and isolation are hard in a highly distributed setting and typically quite expensive
- Why is it important? These guarantees make higher components of a distributed stack much easier to reason about
- What is the approach? Move core piece of concurrency control into datacenter network (i.e., co-design network with software)
- What is the result? Process large class of distributed systems in a single round trip between client to strage without explicit coordination among shards or replicas while providing atomicity, consistency, and fault tolerance with ~10% overhead with up to an order of magnitude higher throughput and ~80% lower latency.
- I’ll be honest…didn’t get the kids to bed until midnight and couldn’t make it past the introduction. In my defense, I’ve seen Dan Ports give a talk on this paper already. See y’all tomorrow.