Keeping Time in Real Systems
summary
Covers concrete tradeoffs: physical clocks drift and UTC inserts leap second repeat that breaks monotonic timestamps, NTP synchronization is slow NTP ≈100ms, Google builds a TrueTime service that exposes uncertainty and forces commits to wait the uncertainty window to guarantee external consistency TrueTime (bounded ±7ms), and high availability systems use logical versions so replicas converge with vector clocks showing causal versus conflicting updates.
AI-written summary · about
chapters (12)
- 0:00Start
- 1:01distributed key-value store
- 2:03consistency model
- 3:50computers have clocks
- 5:40the system clock is a counter kept by hardware, OS kernel
- 8:22The system clock keeps UNIX time
- 9:13interlude: UTC
- 13:26prelude
- 14:57spanner
- 20:05order of transactions == observed order
- 21:37True Time
- 31:29vector clocks
related talks
14mLightning Talk: Beyond NTP: High-accuracy clocks for distributed observability
56mOrdering the chaos - cleaning logs and ordering events in microservices
01h 02mOrdering the chaos - cleaning logs and ordering events in microservices
42mHow to Have your Causality and Wall Clocks, Too
23mDoes It Add Up? Exploring the Delta Temporality in OpenTelemetry & Beyond
35mPredictable Latency
22mLancet: A self-correcting Latency Measuring Tool
21mOptimizing IT Performance in Complex Systems