For failure recovery on a single compute that'll work but doesn't that story get more complex as you add more compute VMs? You know, the actual… ah… balancing of load?
My point was just that the complexity goes up as you want more. At some point, it's worth the money to just use someone else's solution to all of those small problems instead of maintaining your own worse version. Most small teams likely shouldn't be writing their own nginx, haproxy and keepalive configuration for load balancing. The $20-$200/month that it'll cost them is well worth the time that their engineers back.
Of course, there does exist another tier of scale where it becomes questionable to continue paying the premium that comes from asking others to solve your infra problems for you. This falls into the "problems we'd love to be lucky enough to have" bucket for most companies.