What you're saying sounds good on paper, but I don't think real-world experience backs it up.
I've not seen a non-relational system really have sticking power for human knowledge representation yet, despite many people making many attempts at it.
Meanwhile, most SQL databases include excellent support for JSON column types now, if you want to mix in some data that doesn't fit neatly into rows and columns.
And why is that, what do you get out of relational in that case that Solr doesn't give you? The Solr boolean query interface is very impressive and very fast. The caveat is you need a good essentialized schema.
I would call joins a bug of relational rather than a feature. If you have all the data in one collection, you don't need or want joins. For transactions, we do that server side (in our own system) as we have what we call temporary awareness created for each request that only writes when all operations have completed successfully before importing the state changes into primary awareness for that user.
Every large system I've ever encountered that's built on a NoSQL system such as MongoDB or Elasticsearch has inevitably ended up reinventing joins, poorly.
We have 1 collection with 4 doc types (attribute, state, event, process) and 14 relations between them, so a join concept makes no sense for us. I think the schema is the key issue in having success vs not. (Apache Solr - billions of docs)
I've not seen a non-relational system really have sticking power for human knowledge representation yet, despite many people making many attempts at it.
Meanwhile, most SQL databases include excellent support for JSON column types now, if you want to mix in some data that doesn't fit neatly into rows and columns.