Ah but the rub is that it would not be unreliable. One of the things I discovered was that "reliable" in the eyes of non-engineers isn't the same thing. I have

Ah but the rub is that it would not be unreliable. One of the things I discovered was that "reliable" in the eyes of non-engineers isn't the same thing. I have a NAS system for my important data because I know it is reliable, I know dozens of people who "have a copy on a flash stick" because they consider that reliable. You can build "good enough" in a rack that will serve a LOT of clients. One layer of ZFS (1U + 3U) for storage, Five 1U servers, one 2U 10G switch. In my experience that is 1 failure (that results in an outage) every 2 years at most. When I was at Google I (like everyone) had access to failure information about equipment in the "fleet." Motherboards do fail, but VERY rarely, disks are the worst offender but with double parity ZFS you have a week at a minimum to replace a failed drive. ALL the drive failures at Blekko announced themselves before they failed with SMART data. So in addition to the colo cost you've got an engineer spending part of their time watching the rack, fractional head count cost.