[tahoe-dev] How many servers can fail? (was: Tahoe-LAFS Logo selection)

[Nathan Eisenberg]

> To foster my understanding, I've tried to visualize what that means:
> 
>  http://dirk-loss.de/tahoe-lafs_nhk-defaults.png
> 
> Is that interpretation correct?
> 
> Best regards
> Dirk
 
To great extent, yes, although operating with 7 servers in a 3 of 10 encoding is suboptimal.  To keep things simple, you might simply set H=N (10), so that uploads will fail if they cannot be evenly distributed among 10 servers.

Of course, where it gets more complicated is the tree property of tahoe-LAFS - which is to say that files are children of 'directories' which are distributed with the same erasure coding.  To visualize this simply, imagine that you have 1-of-2 encoding on a grid of 4 servers (A,B,C,D).  The directory is stored on (A,B), and the file within it is stored on (C,D).  Should A and B fail, the file (which is NOT stored on A or B) becomes unavailable, even though C and D are online.

I've mentioned previously on the list that this seems to be a design issue, and that perhaps dircaps should be a special type of share that is always distributed (and redistributed) among all tahoe servers regardless of N or K or H.  However, I will readily admit that this is probably easier said than done.

Nathan Eisenberg