Projections in MongoDB

In MongoDB, one should use projections wherever possible. I wanted to quantify two things

We have a table called events, whose stats looks like this: 
db.events.stats()
{
	"ns" : "6ed11a9246404d1b95fe.events",
	"count" : 16912712,
	"size" : 8617342080,
	"avgObjSize" : 509,
	"numExtents" : 25,
	"storageSize" : 9305935856,
	"lastExtentSize" : 2146426864,
	"paddingFactor" : 1,
	"userFlags" : 1,
	"capped" : false,
	"nindexes" : 4,
	"totalIndexSize" : 1167933424,
	"indexSizes" : {
		"_id_" : 548732240,
		"notificationId_1" : 14545104,
		"parameters.notificationId_1" : 54296816,
		"eventName_1" : 550359264
	},
	"ok" : 1
}

So, there are around 17M entries, having a total size of 8.6GB, thus having an average object size of 8.6GB/17M = 509 bytes.

I fetched first three million entries in this collection, using a python program like this: 

class QGMongo(object):
   __conn = None
   @classmethod
   def get_connection(cls):
      if cls.__conn is None:
         cls.__conn = MongoClient('127.0.0.1', 27000)
      return cls.__conn

if __name__ == '__main__':
   conn = QGMongo.get_connection()
   database = conn['6ed11a9246404d1b95fe']
   events = database.events.find()
   count = 0
   for event in events:
      count += 1
      if count == 3000000:
         break

This prgram took around 120s to run.

I set db.setProfilingLevel(2) and then figured out from system.profiles collection that

As per db.system.profile collection,

I then did same experiment using projection: 
class QGMongo(object):
   __conn = None
   @classmethod
   def get_connection(cls):
      if cls.__conn is None:
         cls.__conn = MongoClient('127.0.0.1', 27000)
      return cls.__conn

if __name__ == '__main__':
   conn = QGMongo.get_connection()
   database = conn['6ed11a9246404d1b95fe']
   events = database.events.find({}, {'eventName': 1})
   count = 0
   for event in events:
      count += 1
      if count == 3000000:
         break
This time, the program took around 30s.

As per db.system.profile collection,

As a result of this, I have the answer to my first question: the query times improve significantly using projections. When I used projection such that the data size required by my program dropped to around 15% (1.1GB to 160MB), program running time dropped to 25% (120s to 30s).

However, my second question is not well answered. There are two outstanding questions: