Vlad Horsun wrote:

> If we separate sort record on two parts {key, offset} and
> {not_key_fields} and sort only first part we will do much less work at
> merge stage but final stage will cost more as we will load {not_key_fields}
> in random order from secondary temp file.
>

In others words:
FIRST n on sorted stream and merge joins faster.
Large resultset slower.


Adriano