On Aug 19, 2008, at 12:27 PM, Albert Chern wrote:
Thanks for the heads up. Does anyone know when this change was
introduced?
I am certain that old versions of Hadoop create a new instance of
the key
and value each time a data pair is read.
It was Hadoop-2399. Previously the inner loop had to allocate a new
key and value for each call to reduce. They are now reused, which does
mean that if you need to keep them around, they must be cloned via
WritableUtils.clone. Note that inputs to the map are also reused...
-- Owen