I recall sometime in the mid 2000s there was a fever for achieving five-9s (99.999% uptime, I think -- it became fodder for a few episodes of Mr. Robot). Not that the metric ever went away, but back then a lot of BigIron(TM) vendors advertised achieving five-9s by replacing hardware while the OS remained running and continuing service. Sun 15K and 25K series (Gilfoyle had a used one in the garage running his network) were behemoths whose mem/cpu boards you could swap out wholesale while the entire frame and backplane was powered on, and while the OS the board came out of remains functioning. There were many caveats around the procedure but it worked. Execs and sales guys loved those demos. These monsters were expensive and banks and energy conglomerates were buying them by the dozens. There was also a big todo about hot swappable drives. The idea that you could be doing hardware maintenance while the machine was still running was a novelty, something like brain surgery while the patient was not only awake, but awake and eating, driving his car, talking on the phone, etc.

A decade later I look back with deep surprise that we didn't think to abstract out the service instead of the hardware. I don't know how many of those behemoths are still being bought, now I work almost exclusively with small server instances that can come and go on the fly. Micro services and AWS have taken five-9s in a different direction. I frequently think of Sun as a failed Hephaestus, in a Christopher Nolan film he would be brilliant but could only turn out clumsy tools because of his deformity, he hates the things he makes so he throws them away before completion. Men find these cast-offs and temper and refine them.