In this presentation, I explain how Google’s Spanner system is implemented – and how it’s performance seemingly defies the CAP theorem.
I’ve tried to keep this presentation accessible to people without a deep systems background, but do let me know if you weren’t at my prior presentation about distributed systems. I promise I’ll keep an eye on you throughout today’s presentation.
In this presentation, I:
- Introduce the CAP theorem using a hypothetical groupchat as an example.
- Discuss strict and loose consistency models.
- Explain how Spanner provides consistency without angering the CAP theorem.
- Explain the underlying theory (clock drift error bounds => “wait out the uncertainty”)
- Talk about the expensive hardware that makes this feasible (atomic clocks everywhere!).
- Conclusion: with enough money and hardware, you can create large-scale, strongly-consistent, systems – even when the theoreticians say otherwise.
Slides available here.