The hard-linking phase makes starting up the split shards fast, but the resulting shards contain all the docs from all the other shards too, in a deleted state, but those deleted docs make them a little inefficient at first. Once the shards have been started, a background merge addresses this by rewriting the shard contents in a more efficient form, removing all the deleted docs. While the merge is ongoing it must hold on disk both the old unmerged (hard-linked) data and the new (rewritten) data, plus various temporary files, and in extreme cases this can temporarily need some multiple of the original shard size. Often it doesn't.