I generally agree in regards to shoehorning data into MongoDB. I don't agree that a facebook clone fits appropriately. When thinking through a project, my mental model generally self-configures into third-normal form, which is to say I literally think in the manner of a relational database.
That said, I have found one specific use-case where it has incredibly worked well, which is as a JSON-based file system in which I can query into the document structure to find the records and data I need en-masse with very simple queries. A project I've been working on for a few years has its own custom schema that fits far better into a JSON hierarchy than a relational one. In this specific system there are very few records (compared to our primary RDBMS which has millions), and when not being explicitly maintained, the files are generally static. The single-collection database in question with a few hundred JSON objects would likely have to be split into at least 10 tables with thousands of records each.
When it comes to using said "files" in production, I index the objects into memory for quick retrieval. In order to populate said cache, MongoDB allows me to query into the JSON file structure and retrieve the necessary "files" very simply into however many indices I need. And that's where it has shined in my experience; Not as a live read-write datastore for production use, but as read-heavy storage for a relatively small amount of JSON data with a deep hierarchy that would require an annoying amount of joins and subqueries for otherwise simple queries.
I've also tried it out with our multi-system logging system using capped collections (JSON formatted messages transported and collected with rsyslog). But we grew beyond its limits almost immediately and ended up in far more manageable territory with kibana / elasticsearch.