Disclosure: I work on Google Cloud (and have emailed jh about Discord). Yeah, the most straightforward failure that we can’t migrate away from is the NIC (or r

Disclosure: I work on Google Cloud (and have emailed jh about Discord). Yeah, the most straightforward failure that we can’t migrate away from is the NIC (or rack switch) failing. Obviously, if your path to get off the box is dead, that’s not going to happen :). Others though, like the hypervisor crashing or even host kernel are also possible, but much less frequent than “Hmm, I think the network is dead, we should start a replacement VM”. One of the many motivations for (now) only having Persistent Disk for boot disks, is that it lets us avoid a whole class of truly unrecoverable errors. There are still bad DIMMs, but monitoring for ECC failures often lets us mark the host for replacement in time for the VMs to migrate off before they actually fail.