The 'main' app at my place of work uses MongoDB this way; it even implements a 'relationships' collection that is used for one-to-many 'joins' (ugh, just thinking about it makes me feel ill). Unfortunately, I joined when the initial write was nearing completion and was unable to steer the team in the right direction in time.
I just submitted a proposal for a ground-up rewrite; unless it's accepted I'll be leaving promptly.
In fairness to MongoDB, I've seen this done in the relational world. I once worked at a place who had a schema where everything joined through a single table -- "TableRow_TableRow", which had six fields. Two IDs (which were varchars) and metadata_1 through metadata_4 (also varchars).
They couldn't understand why it was so slow. But hey, it's super flexible, right?
I just submitted a proposal for a ground-up rewrite; unless it's accepted I'll be leaving promptly.