I built one large scale system in my career (when I worked at Google, I made a folding@home screensaver that used up idle cycles in production). When I built i

I built one large scale system in my career (when I worked at Google, I made a folding@home screensaver that used up idle cycles in production). When I built it I ignored 95% of what Google knew about large scale system design because that knowledge was really about building scalable web services and storage systems, while I needed to build a batch scheduler which could handle many millions of tasks simultaneously. We depended on a few core scalable resources available in production (borg, chubby, bigtable, colossus) and tried as hard as possible to avoid spamming them with excessive traffic without adding lots of complex caches and other hard-to-debug systems. In fact, "simplicity" was the primary design goal. The system worked, it scaled, and if I'd followed all the normal Google guidelines, it wouldn't have (because scientific computation and web load balancers differ). Not sure what to take away from that. These days in system design interviews I usually focus on limiting the use cases for the system so that I can architect something that: linear resource consumption for linear workloads over 2-3 orders of magnitude, is simple enough that a small group of engineers can understand the whole system and debug it when it breaks, and not try to optimize for future use cases (clearly documenting the limits of the system) or try to accomodate too many oddball one-off user requests.