Is not reducing the effective cost of a bad update by 10x or more worthwhile? Sure, but if you are rolling out to 1% of users per hour, you detect the problem

Is not reducing the effective cost of a bad update by 10x or more worthwhile? Sure, but if you are rolling out to 1% of users per hour, you detect the problem in a couple of hours and much fewer than 2% of users will have applied the update. This is a relatively small support problem. While if you roll out to everyone at once, you'll detect the problem sooner (within an hour) but have 10x as many affected.