[tahoe-dev] MapReduce over Tahoe-LAFS slides from HadoopWorld

[Kevan Carstensen]

On 10/12/09 7:55 PM David-Sarah Hopwood <david-sarah at jacaranda.org>
wrote:
> ... actually, not even then. I think that servers_of_happiness needs to
> be equal to n for the statement on the slide to be correct -- is that 
> right?

I think so. servers_of_happiness = h guarantees that a successuful
upload will place at least one distinct share on each of at least h
distinct servers. This means that at least h - k (and possibly more [1])
servers can be lost without any data loss. If n = h, then n - k servers
can be lost.

Note that this relation only makes sense for h >= k, since h < k would
yield negative servers. That case is probably best interpreted (along
with h = k) as not assuring any integrity in the event of server failure.

[1] The "and possibly more" comes from the rather loose coupling between
the peer selection process and the servers_of_happiness behavior. If
possible, tahoe-LAFS will attempt to place each of n shares on its own
distinct server -- if this succeeds in the case where h is smaller than
n, then it is possible to lose more than h servers without losing the file.
-- 
Kevan Carstensen | <kevan at isnotajoke.com>