Steps to Reproduce Issue :

1.) Have a fairly large collection, ~10 million records, avg doc size is 1.4 kb.
2.) Replace each document (all 10 million) frequently with a remove() -> insert() (This works faster than an update because each remove+insert can be batched)

After some time, according to stats the StorageSize gets huge, even though the totalSize is relatively small. In my case the StorageSize is 17 times larger than the totalSize.

A single shard would have a 5 Gb total size, but a 100 Gb Storage Size.

Here is the stats() for the collection in question

http://pastebin.com/qsbdJMQV

RS4 and RS5 are relatively empty because they are new shards.