HTTP traffic is different from run this CPU intensive pipeline service for 10 minutes and now increase the volume by 100x in a few months.
I used to manage a popular app/website that had tens of billions of real user page views per month. It wasn't easy but it also wasn't hard. It was surprisingly easy with a single large Postgres DB, a lot of caching, optimized tables and queries.
Yes, but their shapes are very different. Most users interact with a small, geographically constrained bubble. It is (comparably) an easier problem for the usage patterns of Facebook.
The app is also designed so when you open Insta, only a few images are shown, infinite scroll is there for a reason. The search page suggestions? Probably from a cache and then re-ranked quickly.
GitHub sells this as a service.