consider chainset replication

[thraxil]

as described here: http://hackingdistributed.com/2014/02/14/chainsets/

for large cluster sizes, it reduces the chance of data loss. Need to evaluate whether the additional overhead would be too much on smaller clusters though.