Currently, we use BSONObjBuilder to create BSON then parse it into the form we need during $group optimization time; this happens in several places here. It would be better if we could avoid the step of going to BSON first; this could prove more costly than necessary for multiple $groups/$groups with a large number of fields.