This is an interesting concept. Are there any relevant studies/articles of this in practice, beyond the Anecdotal?
It seems to me that it is hard to test whether the random rollout just aims at preventing server crashes or it has behavioural connotations as well.
A slow roll out does allow you to catch the inevitable 1.0 bugs before too much damage is done.
Also, at least in the past, FB used to roll out redesigns to groups that were relatively isolated from the rest of the graph like New Zealand.