As a reminder, see the discussions in PRs #604 and #585; We could likely speed up things by consuming the result cursor into a map and then assigning values to requested keys by map look-ups. For larger number of keys - with the threshold to be defined - this should be faster than the nested looping. The idea is to use one or the other approach depending on the number of requested keys.

This all is assuming we cannot find a way to get back the results in the order we expect.