No, of course they need to have very good reliability. The problem is the old curse of the redundancy: to make things redundant, you need a system which integrates the childs. For example, in a hw raid array, you need a raid controller. If this controller dies, your whole array dies, although it was redundant.
Or if you try to make redundant servers, you need something which finds out which server is unavailable and redirects traffic to its clones. This thing can also die.
There are completely redundant systems possible, but they are normally too complex, and most sysadmin simply won't do that. Most of them simply don't even understand the concept. In most cases, they are using not redundant, but very reliable systems to connect and manage the redundant cluster of not so reliable sub-systems.
For example, there is a very costly raid controller, using a redundant disk array from the far east.
It is nearly always the case, but this is which won't work for UPSes. Why?
- UPSes are working with current. If the primary power source is lost, the secondary need to be powered on in 1/100 - 1/10000 seconds, or the outer voltage of the internal transformators of the servers will significantly drop.
- UPSes can't be tested. You can't plug out the power to find out if an UPS is okay, because your server will be shut down if it isn't.
For real ups redundancy you needed to use a redundant server cluster using servers with redundant power sources.
In most case, in big server farms, it is not done. Yes, google will also die if the actual server node dies which you are actually using. But they will find the bad nodes very fast, and eliminate that from their system.