Yeah, this is a valuable question, and I agree its not an obvious answer. The tricky part is that at enough scale, humans are really the bottleneck. Teams step on each others toes, abuse schemas, add data to a schema "because", etc, etc. Its possible that to best manage the humans, de-normalization and relying on replicated data stores with slightly different views of the data is simplest.
Add again if "overly flexible" is a long term product requirement, I'd argue you're going to eventually need full text search with all the power of Lucene (I'm betting its on your roadmap).
If Fauna perfectly addresses this problem domain, its likely quite helpful, but this article did not convince me it'll always be cheaper / better than DynamoDB + ElasticSearch for the complex usecases. That said, I look forward to the day I'm proved wrong :)
