Amazon strongly implies that your data is stored triple redundantly. You'll have to buy more than one disk to get triple redundancy. Also, you'll need to "scrub" the data regularly to detect bit rot (100 bytes per TB are expected to go bad every year). And probably store the files with redundant coding to protect against this bit rot.
Plug: I am working on a startup (submitted to this YC round!) to solve this problem using user^Wcustomer-owned hardware for precisely the sort of reasons you describe. I am looking for co-founders. If anybody wants to talk, email is in my profile.
One external online provider really only counts as "one copy", ever. This is primarily because you cannot audit the ongoing storage architecture and processes of any given provider. You're looking for SPOFs, not how many disks may hold data replicas. One software error (or site/account hack) can wipe out all of your data. Or an entirely out-of-band error occurs: the provider goes belly-up.
Cloud storage is awesome in many ways. Yet it doesn't replace your backup strategy, it merely complements it.
'scrubbing' is just a fancy way of saying that the files are read from media, checksums recomputed, and compared against stored checksums. If checksums are different and data was stored redundantly, then recovery is carried out, and correct data is written back to media.
I'm too lazy and don't really do it with my multiple DVD, CDs and HDD backup directories. But ideally, I should be doing it. My startup will make this sort of thing easy and automatic.
Search for zfec. Allows you to split a file into N chunks, only M of which are required to reconstitute the original. Protects against N-M independent errors/corruptions.
Sorry, I think I should retract that statement which I seem to have recalled mistakenly. The error rate seems to be quite a bit lower than that, so I will post an article here after I research it thoroughly.
Sorry, my comment above wasn't well thought out. Amazon's durability guarantee is on a per-object basis, not on a per-byte basis. I will post an article here after I research it thoroughly.
Plug: I am working on a startup (submitted to this YC round!) to solve this problem using user^Wcustomer-owned hardware for precisely the sort of reasons you describe. I am looking for co-founders. If anybody wants to talk, email is in my profile.
Plug #2: I wrote and submitted this article about Glacier yesterday but it sank fast: http://psranga.github.com/articles/possible-architecture-of-... Email me if you want to talk about going up against an 800 lb gorilla. :)