The Importance of Variance in Distributed Systems

Besselos: I think that we need to start taking variance more seriously as a core property of our distributed systems. We've built some analytics that will do analysis of performance distributions, and we have found that these are good indicators of customer surprise. And it turns out that if you were able to break up your workload into pieces that are kind of self-similar, then many of these behave consistently over time. So figuring out how to measure those and measure those at fine granularity ends up being a big part of the challenge here.