One of the propositions of cloud is that it should be possible – through the use of intelligent software – to build reliable systems on top of unreliable hardware. Just like you can build reliable and affordable storage systems using RAID (Redundant Arrays of Inexpensive Disks).
One of the largest cloud providers says: “everything that can go wrong, will go wrong”.
So the hardware is unreliable, right? Mmm, no. Nowadays most large cloud providers buy very reliable simpler (purpose-optimized) equipment upstream of suppliers in the server market. Sorry Dell, HP & Lenovo there goes a large part of your market. Because when running several hundred thousands of servers a failure rate of 1 PPM versus 2 PPM (parts per million) makes a huge difference.
Up-time is further increased by thinking carefully about what exactly is important for reliability. For example: one of the big providers routinely removes the overload protection from its transformers. They prefer that occasionally a transformer costing a few thousand dollars breaks down, to regularly having whole isles loose power because a transformer manufacturer was worried about possible warranty claims.
The real question continues to be what happens to your application when something like this happens. Does it simply remain operational, does it gracefully decline to a slightly simpler, slightly slower but still usable version of itself, or does it just crash and burn? And for how long?
The cloud is not about technology or hardware, it’s about mindset and the application architecture.