Suppose you have to roll out a change to 500 machines. Maybe it is a new kernel. Maybe it is just a small bug-fix.
Do you roll it out to all 500? No. You roll it out to a small number of machines and test to see if there are problems. No problems? Roll out to more machines. Then more and more until you are done.
These early machines are called "canaries".
The classic example of animals serving as sentinels is the canary in the coal mine. Well into the 20th century, coal miners in the United Kingdom and the United States brought canaries into coal mines as an early-warning signal for toxic gases including methane and carbon monoxide. The birds, being more sensitive, would become sick before the miners, who would then have a chance to escape or put on protective respirators. Source: Wikipedia
Here are some canary techniques:
Do one machine (maybe your own desktop), do some machines (maybe your co-workers), do many machines (larger and larger groups until done.) Any single failure means you stop the upgrade, roll back the change, and don't continue until the problem is fixed.
Upgrade 1 machine, then 1% of all machines, then 1 machine per second until all machines are done. (Typical at Google and sites with large clusters)
This procedure can be done manually but if you use a configuration management system, the ability to do canaries should be "baked in" to the system.