The hitchhiker's guide to distributed transactions
summary
Surveys systems that reduce commit latency from two WAN round trips to one WAN round-trip, comparing concrete mechanisms: leader-based consensus (Spanner/pipeline) versus leaderless or fast paths (MDCC, Carousel using fast Paxos), replicated-commit across full-DC copies, deterministic-function execution (OceanVista/SLOG), and TAPIR’s inconsistent replication. Shows trade-offs: isolation falls from serializable toward read committed with generalized Paxos, gateway-to-leader hops vs direct replica proposals, and implementation costs (parallel commits cut commit latency but inflate transaction records); overall achieves one-RT commits with different safety and read-cost trade-offs.
AI-written summary · about
chapters (15)
- 0:00Start
- 6:49The Hitchhiker's Guide to Distributed Transactions
- 12:53Replication and Fault Tolerance
- 18:45Basics of Transactions
- 19:47Effects of Concurrent Transactions on One another
- 20:47Isolation Levels
- 25:52Multi-Node Transactions
- 28:35Multi-Shard Transaction
- 32:12Pipeline Transactions
- 36:18Parallel Commits
- 49:33Transaction Model
- 58:54Replicating the Transaction
- 1:05:03Inconsistent Replication Scheme
- 1:06:25Summary
- 1:13:58Q a Session
related talks
37mGeo-Replicated Transactions in 1.5RTT
01h 00mDistributed transactions implementation trade-offs
51mDo Not Reimplement Distributed Consensus Algorithms
30mThe Architecture of Microservice Transaction Based on Distributed Consensus with Kubernetes
01h 02mConsistency and Agreements in Distributed Systems
28mPractical Approaches To Scaling Distributed Transactions
41mTransactions: myths, surprises and opportunities
48mDistributed Consensus Algorithms