I've always used Postgres. For a medium-sized app that I work on right now it's running great. Were I to seriously need more throughput I'd either shard (no small task I know) or use CockroachDB (which to my understanding is basically Postgres with built-in sharding and no extension support). Throwing away relationships, constraints, unique compound indices and all of the tools I love that Postgres provides just to get schemaless JSON with high write throughput out of the box doesn't sound like a good deal. But so many people have made the decision to go NoSQL so there must be something I'm missing.