It's somewhat counter-intuitive, but it's actually documented in the API docs -- Hadoop reuses the keys / values, you should clone them if you want to keep them around.