The vast majority of IPC already used sockets decades ago, so it was easy to split a composition of services across multiple machines (just change the config fr

The vast majority of IPC already used sockets decades ago, so it was easy to split a composition of services across multiple machines (just change the config from unix domain- to network sockets). Anything that ran commands to do things could be wrapped in ssh (for example, that's how I offloaded image processing for a sharing site). Two layers of DNS records solve the addressing issue: each box gets a static codename with an A record and each service gets a descriptive name with a CNAME record pointing at the box it's on. When you move a service, change the CNAME to point to the new box. If you have full box replicas/backups, you can also switch the box name or reroute at the IP level. Preferably you're running your own DNS and can issue a flush command to your entire fleet so you're not stuck waiting for propagation.