On Monday, May 4, 2020 3:19 AM, antlists <antli...@youngman.org.uk> wrote:
> On 03/05/2020 22:46, Caveman Al Toraboran wrote: > > > On Sunday, May 3, 2020 6:27 PM, Jack ostrof...@users.sourceforge.net wrote: > > curious. how do people look at --layout=n2 in the > > storage industry? e.g. do they ignore the > > optimistic case where 2 disk failures can be > > recovered, and only assume that it protects for 1 > > disk failure? > > You CANNOT afford to be optimistic ... Murphy's law says you will lose > the wrong second disk. so i guess your answer is: "yes, the industry ignores the existence of optimistic cases". if that's true, then the industry is wrong, must learn the following: 1. don't bet that your data's survival is lingering on luck (you agree with this i know). 2. don't ignore statistics that reveal the fact that lucky cases exist. (1) and (2) are not mutually exclusive, and murfphy's law would suggest to not ignore (2). becuase, if you ignore (2), you'll end up adopting a 5-disk RAID10 instead of the superior 6-disk RAID10 and end up being less lucky in practice. don't rely on lucks, but why deny good luck to come to you when it might? --- two different things. > > i see why gambling is not worth it here, but at > > the same time, i see no reason to ignore reality > > (that a 2 disk failure can be saved). > > Don't ignore that some 2-disk failures CAN'T be saved ... yeah, i'm not. i'm just not ignoring that 2-disk failure might get saved. you know... it's better to have a lil window where some good luck may chime in than banning good luck. > Don't forget, if you have a spare disk, the repair window is the length > of time it takes to fail-over ... yup. just trying to not rely on good luck that a spare is available. e.g. considering for the case that no space is there. > > this site [2] says that 76% of seagate disks fail > > per year (:D). and since disks fail independent > > of each other mostly, then, the probabilty of > > having 2 disks fail in a year is: > > 76% seems incredibly high. And no, disks do not fail independently of > each other. If you buy a bunch of identical disks, at the same time, and > stick them all in the same raid array, the chances of them all wearing > out at the same time are rather higher than random chance would suggest. i know. i had this as a note, but then removed it. anyway, some nitpics: 1. dependence != correlation. you mean correlation, not dependence. disk failure is correlated if they are baught together, but other disks don't cause the failure (unless from things like heat from other disks, or repair stress because of other disk failing). 2. i followed the extreme case where a person got his disks purchased at a random time, so that he was maximally lucky in that his disks didn't synchronize. why? (i) offers a better pessimistic result. now we know that this probability is actually lower than reality, which means that we know that the 3.5k bucks is actually even lower. this should scare us more (hence us relying on less luck). (ii) makes calculation easier.