To add to Todd/Ted's wise words, the Hadoop (and MapReduce) architects
didn't impose this limitation just for fun, it is very core to enabling
Hadoop to be as reliable as it is. If the reducer starts processing
mapper output immediately and a specific mapper fails then the reducer
would have to know how to undo the specific pieces of work related to
the failed mapper, not trivial at all. That said, the combiners do
achieve a bit of that for you, as they start working immediately on the
map out, but on a per-mapper basis (not global), so easy to handle
failure in that case (you just redo that mapper and the combining for it).
-- amr
Ted Dunning wrote:
I would consider this to be a very delicate optimization with little utility
in the real world. It is very, very rare to reliably know how many records
the reducer will see. Getting this wrong would be a disaster. Getting it
right would be very difficult in almost all cases.
Moreover, this assumption is baked all through the map-reduce design and
thus doing a change to allow reduce to go ahead is likely to be really
tricky (not that I know this for a fact).
On Mon, Jul 6, 2009 at 11:14 AM, Naresh Rapolu <nareshreddy.rap...@gmail.com
wrote:
My aim is to make the reduce move ahead with reduction as and when it gets
the data required, instead of waiting for all the maps to complete. If it
knows how many records it needs and compares it with number of records it
has got until now, it can move on once they become equal without waiting
for all the maps to finish.
