When bulk building column indexes we can take advantage of the fact that a collection scan produces results in RID order. (We should confirm this with the storage exec team)

As we walk the collection, we maintain a table from (path -> [list of cells ordered by RID]). For each document we take out of the collection, we produce all of the cells for it, and append them to the corresponding list.

At the end, we sort by the path (the key of our table) and then insert each list into the index. This avoids doing a blocking sort over the entire set of keys.