Centralized management is very useful, just a random delay is not enough. One of the (big) companies I worked with had jury rigged something with chef I believe to show different machines different "repositories" and roll things out progressively (1% of the fleet, 5%...).