Earlier quoted context omitted.
Scott Hanselman had an interview with the Stack Overflow Director of Engineering just a few weeks ago. They cover some high-level details there [0]. As I understand it's an .NET Core monolith with a very-large (~1.5 TB RAM) SQL Server backend. [0] https://hanselminutes.com/847/engineering-stack-overflow-wit...
One thing about using SQL databases live in production for everything is that the architecture is amazingly simple, almost always performant, but if something goes wrong, it goes catastrophically wrong. Amazon for new critical applications prefers NoSQL solutions for this reason since performance is flatter and not dependent on vagaries of the query optimizer.
IMHO the logical progression for applications with typical relational models is starting with an RDBMS in combination with an ORM, breaking out of the ORM when necessary for more complex/slower queries, ditching the ORM altogether when it starts to break, then evaluating sharding or moving domains into a NoSQL solution piecemeal. A well-architected monolith or series of microservices suits this paradigm which allows you to realize the productivity benefits of SQL and only worry about the deduplication, indexing, and consistency issues that come with NoSQL when it becomes actually necessary.
Then again if I were Amazon I'd probably plan for scale up front and be jaded by their experience with Oracle too.