[Openstack] [Swift] Erasure code durability and overhead in small clusters
On 14/03/18 13:33, Mark Kirkwood wrote:
> Doing a bit more playing about leads me to think that for a *single
> region* EC policy we can get a tighter lower bound on the number of
> hosts: I'm calculating it as (k+m)/m.
I probably should have shared the reasoning rather than just plumping
the formula out! So here it is:
Suppose we have n hosts. The k + m fragments get spread about these n
hosts,. On average, therefore (k + m)/n will be on any one of them. We
want to ensure that this number is <= m (so one losing 1 host does not
take out too many fragments). i.e:
(k + m)/n <= m
Doing schoolboy algebra on this gets:
(k + m)/m <= n
Which is the stated conjecture for the minimum number of hosts.