M320: Subset Pattern versus Projections

I am interested in hearing a contrast between the Subset Pattern (separating information and information details into separate collections using a 1:1 relationship) and using projections (combining all into an aggregate collection and having queries work with the fields of interest) and using specific field updates โ€“ versus working with the entire document.
When using projections and field update commands โ€“ is the underlying document brought into memory anyway?

Using projections is a good practice.
However, even if you decide to project few fields out of a one megabytes document, the server would still bring the whole document (1MB) in RAM.
The savings from the projection happen because only the few fields will be returned to the client through the connection.