If you want to write a truly scalable application you structure everything such that you do joins in your application layer.
http://highscalability.com/ebay-architecture
And in the case of MongoDB you avoid joins since it is a document database. You embed data instead.
Most applications fit very well in a single server. For example, Stack Overflow runs on a single instance of SQL Server, replicated to a slave in another data center. In such a case, the convenience of joins and transactions is priceless.
And even at scale, it makes sense to rely on joins and transactions. The perfect example is AdWords that runs of F1 and Spanner:
"Our users needed complex queries and joins, which meant they had to carefully shard their data, and resharding data without breaking applications was challenging."
http://static.googleusercontent.com/media/research.google.co...
And in this situation the rule of thumb is to do joins in your application layer so you can store different types data (e.g. graph/document/relational/unstructured) in different systems and easily cache where needed.
The fact that so many new databases have been created in the last decade suggests that there are a lot of people who do fall into this camp.
They have about 7.5m visits per day to Stack Overflow. That's about 86 requests per second, which perhaps at peak is several multiples larger than the average. I wouldn't consider that a particularly gargantuan website. That sounds like traffic that you could service with a reasonable fleet of web servers and database fleet, given caching. Quite a lot of actions on the site can be applied with eventual consistency, I'd imagine.
http://stackexchange.com/sites?view=list#traffic
I don't mean to say that the system or the problem that it's solving is trivial - I am sure it is difficult to get right. As websites go, it might be large, but as systems go it's not particularly high traffic among systems that receive traffic from machines rather than humans. Imagine that you operate a data center, and you want to sample CPU, memory, etc. from your machines every minute. If you collect 50 samples per machine per minute, and you have 258 machines, you'll be handling 86 samples per second. Storing 86 samples per second into a time series database is probably considerably easier than SO's website rendering, but it goes to show that high-traffic or high-frequency systems are common in companies beyond small to medium size. It is easy for cross-cutting concerns like this to have massive request volumes, far greater than the human-generated traffic to any website.
To anyone else reading: no, your use case probably isn't so special that the solution Stack Overflow arrived at just doesn't work for you.
I will be happy if i got such simple tasks :)