In iterating over the cursors in a table we manually create a new BSONObjBuilder for each sub. This will induce a new allocation for each sub. If we can rework the code to concatenate all of the {{sub}}s and their children and then feed into the parent BSONObjBuilder supplied in the function, we can avoid these individual allocations.

I have a patch POC that clumsily achieves this, but we'd need somebody to make sure the final approach is sound.