I do this with cloud VMs these days: it's particularly useful when there's third party code in the mix. The theory is that your uptime represents a known tested amount of operation (roughly), and as such everytime you go beyond that in production but not testing, you're getting into unknown space - systems are too complicated to validate everything (see the classic patriot missile system bug [1]).

So, you should never deploy anything to production which won't be rebooted more frequently then you reboot it in testing. In practice, you should probably reboot much more frequently - as frequently as possible - to keep the delta between "known good" and "mutated" as small as possible.

[1] https://en.wikipedia.org/wiki/MIM-104_Patriot#Failure_at_Dha...