flamegraph.svg

On my workstation, running

db.test.aggregate([{$group: {_id: null, total: {$sum: "$a"}}}]) 

on 16 million {a:2} documents takes about 10 seconds. The attached flamegraph shows that roughly 4% of the perf samples collected occur in the hash function for HashAggStage's _ht hash table.

Given this query is guaranteed to return only one document, there is no need for the hash table in the aggregation. It probably makes sense to have a new plan stage, to separate the code from HashAggStage and to make explain output clear.

Getting the collection filled with documents can take some time, I found this python snippet to be helpful

for (let i = 0; i < 1000*4; i++) { db.bar.insertMany([' + ','.join(['{a:2}' for _ in range(256)]) + "])}