My hunch: a more invasive one. Think of turning off all machines in a cluster for major power work or to replace the enclosures themselves. Maintenance on a sin

My hunch: a more invasive one. Think of turning off all machines in a cluster for major power work or to replace the enclosures themselves. Maintenance on a single machine or rack, instead, happens all the time and requires little more scheduling work than what you do in Kubernetes when you drain a node or a group of nodes. I used to have my share of "fun" at Google sometimes when clusters came back unclean from major maintenance. That usually had no customer-facing impact, because traffic had been routed somewhere else the entire time.