Anything that requires a reboot, for example, is likely to be held off for a long time. When people are at the computer, it's because they want to use it right now. They aren't there to watch it reboot a few times.

It's less painful once you scale past single systems. Insofar as you have distributed systems that tolerate some fraction of their instances disappearing and reappearing, you can roll an upgrade out across the system without users having to notice that it's happened. Where I work we do this very well -- we update and upgrade live production systems for CVEs and feature adjustments quite frequently without systems running on them being any the wiser.

Still remaining is the sheer amount of independent dependency tracks into any given software system. Operating system, multiple language ecosystems, multiple infrastructure systems ... once you include development dependencies and tools -- and you ought to -- the surface area to watch and manage goes up sharply.