Echoing Michael Segel, why not subclass and set reducers to whatever you
want in your subclass?
But you probably don't want to have reducers anyways. The output from your
mappers will have to be sorted and fed to the reducers which will put up a
load on your cluster, a loading that could be better deployed moving the
data to S3.
Or limit the number of mappers you have running at any one time via
configuration or in a subclass limit the rate at which they write?