Google Spanner is consistent across continents. The trick: it waits.

transcript

Google Spanner is strongly consistent across continents. It doesn't beat the trade-off. It pays for it, a few milliseconds at a time.

Here's the trade-off. You want instant visibility, and high availability. And low latency. You can't have all three at once. Want the first two? Then you wait. You give up latency.

Spanner's secret is TrueTime. GPS receivers and atomic clocks, in every datacenter. Ask it the time, and it doesn't give one answer. It gives a range. Earliest to latest. That uncertainty is typically under 7 milliseconds.

Now a write. It gets a timestamp. Then, before committing, it waits out that uncertainty. On purpose. Then it commits. Every datacenter agrees on the order of writes.

The price? Writes take about 10 to 20 milliseconds. A write commits once a majority of Paxos replicas agree. Across regions, 100 milliseconds or more. And it still targets 99 point 9 9 9 percent availability, multi-region.

No eventually consistent surprises. Ever. Spanner pays in write latency, never in stale reads.

Full breakdown in the lecture.