See here: http://github.com/elasticsearch/elasticsearch/issues/issue/77.
Basically allows to continue and scroll a search request quite easily, with
faster execution then if executing the search again. It comes at the expense
of resources used to maintain the scrolling state on each node.
Note also, that scrolling is in a point in time (the time the first search
was executed) scrolling. Any changes applied to the index, even with a
refresh, will not be taken into account when scrolling (which is what you
would want). This guarantees that unique results will be returned. Think of
it as opening a cursor in database lingo.