Presumably, most updates to system interfaces are designed to achieve some goal related to the owner of the system. Often this would be usability, but of course, it could be something else like profitability, security, etc.
Interface changes for non-usability related goal:
So the first point is that a subset of interface updates are performed with a goal other than keeping the user happy. Take for example, news sites that split articles over multiple pages. This generally reduces usability, but increases ad impressions. Another update on stackexchange meant that the stack exchange no longer works with the Reddit toolbar. This was done to prevent other websites taking advantage of stackexchange content through framing. The fact that it had collateral negative effects on usability was accepted.
Interface changes that fail to increase usability: Sometimes a change is introduced that is notionally designed to improve the interface of a system, but it fails to do so. Sometimes this is a failure in general. Sometimes this is just a failure for some individuals. For example, when websites introduce a flash landing page in the hope of improving the site, but they actually substantially reduce the usability.
Usability increased but disruptive: Another class of usability updates are related to the disruption caused to existing users workflows. Automaticity is an important part of skilled performance, both in terms of the speed and the low demands it places on limited attentional resources. The more time people spend with a system, the more they automate the steps required to use the system to achieve their goals. When a system changes, users are often forced to learn a new way of doing things. Even if this new way is better in a general sense, the disruption to automaticity can induce frustration and a drop in performance to existing users.
I remember seeing something similar when I started using OSX as my primary operating system. It was quite disruptive to my workflow, because for many activities that I'd automated on Linux and Windows systems (e.g., application switching, keyboard shortcuts, text editing, arranging windows, accessing menus, etc.) the workflow was a little bit different. I remember reading an article by a Windows user who listed 65 reasons why Macs suck. The user was trying to apply their Windows workflow to an OSX workflow and was feeling frustration because it didn't always work. More importantly, they hadn't grokked the new workflow required. Grokking the new workflow takes time as new strategies are discovered and automated (this was my attempt to respond to the critique). It also takes time for the implicit learning processes to gradually learn the relative speed and cognitive demands of the various strategies.