Hacker News Reader

Comment by theandrewbailey | Hacker News Reader

> Imagine tossing a coin that has a 0.001:99.999 probability split. Just 0.001% chance that you will be killed vs 99+% chance that you won't. Believe me, you don't want to be tossing this coin in a macro scale OR way too many times.

That has to be some kind of named law: As the volume of a system increases, tolerance for failure decreases. Used in cases of: 1% failure of a hundred might be acceptable, but 1% failure of a million isn't.

View on HN

Weirdly enough, we see this all the time in large systems. When we built DynamoDB - the core 3-way replication is based on the fact that machine failure rates are something like 0.005, so we will need to be unlucky (0.005)^2 times. Guess what? it happens alarmingly often enough when you run 100k boxes.

Just call it "Murphys Law of Large Numbers":

The higher your scale, the faster it will fail.

We (humans) run on some 37 Trillion cells and don’t fail very often compared to computers; blue whales scaled for size could have ten quadrillion cells and they don’t get cancer.

Maybe its not a law of large numbers, but one of medium numbers (in the sense of Dawkin’s Middle World is human scale or Ron Graham speculating that maybe all the most interesting math is in the very large numbers we can’t work with)

We fail all the time. We get illover every silly thing, develop cancer, break down little by little over time. It's just that we evolved to handle this better than a artificial designed high spec-machine which was primary optimised for low budget.

Reply on news.ycombinator.com