Out of curiosity, what would happen if the 3 nodes that contain the particular copy of a file all failed? Wouldn't this mean that the system would no longer have any access to that particular piece of data?

rmanne

@totofufu the probability of 3 nodes failing before one of them is reinstall is ridiculously low. The probability of 1 node failing is already extremely low (less than 1/1000000000) and now cube it (less than $$1/10^{27}$$). Yes, if all 3 failed, the files are lost. This is true for any type of persistent system, like RAID, AFS and even AWS. Imagine if there was some massive explosive earthquake or something that somehow destroyed every single one of the servers that hosts AFS. Well of course you can't recover from it, but luckily, the probability of that happening is extremely low.

In other news, this is true of a lot of primality tests as well. The good polynomial ones (there is one apparantly in n^6, but your computation is basically never going to complete right...) are also probabilistic, but run them 50 or more iterations and the probability that you have a false positive is less than $$1/2^{50}$$.

totofufu

@rmanne wasn't aware that the probability of a node failing was that tiny! Thought it was like a weekly occurrence, thanks for clarifying that up :]

rmanne

@totofufu that's the probability of a node failing and you not being able to replace it and make the node working again before the next one fails. The exact probability of failure depends on the hardware used, but the probability 3 fail at the same time is extremely low.

Out of curiosity, what would happen if the 3 nodes that contain the particular copy of a file all failed? Wouldn't this mean that the system would no longer have any access to that particular piece of data?

@totofufu the probability of 3 nodes failing before one of them is reinstall is ridiculously low. The probability of 1 node failing is already extremely low (less than 1/1000000000) and now cube it (less than $$1/10^{27}$$). Yes, if all 3 failed, the files are lost. This is true for any type of persistent system, like RAID, AFS and even AWS. Imagine if there was some massive explosive earthquake or something that somehow destroyed every single one of the servers that hosts AFS. Well of course you can't recover from it, but luckily, the probability of that happening is extremely low.

In other news, this is true of a lot of primality tests as well. The good polynomial ones (there is one apparantly in n^6, but your computation is basically never going to complete right...) are also probabilistic, but run them 50 or more iterations and the probability that you have a false positive is less than $$1/2^{50}$$.

@rmanne wasn't aware that the probability of a node failing was that tiny! Thought it was like a weekly occurrence, thanks for clarifying that up :]

@totofufu that's the probability of a node failing and you not being able to replace it and make the node working again before the next one fails. The exact probability of failure depends on the hardware used, but the probability 3 fail at the same time is extremely low.