You design the complete system so it does not have dependencies on a single component, that way each subsystem can have proper maintenance schedules. It takes a lot more up-front work, along with defining what levels of service degradation are acceptable during a window, but that's the difference between a professionally run service and one run by amateurs. Look up Google's SRE methodology and their concept of error budget.