Re: [Gnu-arch-users] Re: [GNU-arch-dev] Re: [ANNOUNCEMENT] /Arch/ embraces `git'

[Jacob Gorm Hansen]

Matthieu Moy wrote:
> John A Meinel <address@hidden> writes:
>
>
> > However, I think 2^-160 * 2^-10 is a low enough probability of collision
> > that we don't have to worry about it. (Even if it is only 2^-80).

With 2^80 files the probability is 50%, as per the birthday paradox, but 
that is not the point. The point is that a) hash functions were not 
designed for this, so unless there is some clear strategy for handling 
collisions some users _will_ see silent corruption of their archives, 
and b) the whole git approach of storing full copies of everything and 
needing 160+-bit ids to index things is just gross, and in the end it 
buys you no extra functionality, it just wastes a lot of disk space for 
no apparent reason.  Linus Torvalds needs to  show the world that he was 
not an idiot for going with bk in the first place, git is his panicky 
'solution', and right now it seems all the major free SCS's are bending 
over backwards just to get his attention.

> I guess the probability of collision is much lower than the
> probability of hardware failure (like a 0 turning into a 1 in the RAM
> of your computer).

That is a problem that can be fixed at the hardware-level, i.e. with 
error-correcting memory. For instance, IBM has been selling 'chipkill' 
memory with their top servers for several years. Using this argument to 
justify hash collisions is like trying to convince the cop who just 
pulled you over for speeding that he should go and chase some 'real 
criminals' instead.

Jacob