Geo-Replicated Transactions in 1.5RTT
summary
A system called Consus implements a new commit protocol that claims to commit fully general geo‑replicated transactions in three wide‑area message delays (an explicit 1.5RTT claim) without a central coordinator, using generalized Paxos (journalized Paxos) and per‑DC replicated state machines. Author builds ~32k lines of code, runs TPCC on EC2 with artificial RTTs, and measures latency near the predicted 1.5RTT; stock‑level transactions show better-than-expected behavior. Caveats: GC, deadlock handling, and the implementation are functional but not production-hardened.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:53Geo-Replication: Primary Backup
- 6:46Stored procedures and one-shot transactions
- 10:16Commit Protocol Assumptions
- 10:57Commit Protocol Basics
- 11:59Overview of the Commit Protocol
- 13:06Observing vs. Learning Execution Outcomes
- 14:29Traditional Paxos
- 17:03Generalized Paxos Fast Path
- 18:13Generalized Paxos Example
- 21:12Cauterizing Loose Ends
- 27:02Paxos All the Things!
- 32:14Implementation
- 32:41Evaluation Setup
- 34:58TPC-C Stock Level Latency
- 35:45Summary
related talks
01h 14mThe hitchhiker's guide to distributed transactions
20mExploiting Commutativity For Practical Fast Replication
17mMako: Speculative Distributed Transactions with Geo-Replication
01h 12mSimplifying global-scale strong consistency
54mEvolving Transactions in Apache Cassandra
01h 00mDistributed transactions implementation trade-offs
41mTransactions: myths, surprises and opportunities
47mBeing transactional in 2019 - what does this actually mean?