Keynote: Building Spanner
summary
Spanner is a global relational storage system built over Paxos replication and a novel TrueTime clock that provides bounded absolute error, developed five years and in production for a year with a revenue‑critical migration from sharded MySQL (F1). It assigns commit timestamps from TrueTime GPS+atomic oscillators and waits until that timestamp is definitely in the past, combining Paxos replication with a hybrid 2PL + snapshot isolation concurrency model to offer serializable, globally consistent snapshot reads with typical epsilon of 1–7 ms.
AI-written summary · about
chapters (12)
- 0:00Start
- 6:28Megastores Shortcomings
- 9:15Business Concerns
- 9:41Simplified Diagram of Spanners Data Model
- 14:35Interlude Slide
- 24:54True Time
- 47:51Structured Keys
- 48:25Protocol Buffers
- 49:27Relational Data Model
- 49:59The Sequel Engine
- 55:21How Do You Prove There's no Dependencies
- 1:06:49Concurrent Data Structures
related talks
49mSpanner - fully managed horizontally scalable relational DB with ACID transactions
41mSpanner - Google's Distributed Database
22mOrder And Consistency At Scale
47mGoogle Cloud Spanner: Global consistency at scale
33mGoogle Cloud Spanner: Global consistency at scale
37mGoogle Cloud Spanner: Going global with relational database
53mKeynote - Jepsen 7: Anna Concurrenina
12mSpanDB: A Fast, Cost-Effective LSM-tree Based KV Store on Hybrid Storage