Should it be used as a database in the way that Postgres would be used as a database? It seems pretty unfit for that role, with its past durability issues and that writes have to be tailored to how the data will be consumed, due to lack of ad-hoc queries.

As something layered on top of the actual source of truth it may be reasonable, writing a materialized version of a costly query that's read very often, sure, but that's an optimization and it'd be competing with Redis for what matters there.

Same thoughts. I see mongo as a key-value store (string key -> string value), with a json wrapper over the values.

MongoDB has pipeline queries https://www.mongodb.com/docs/manual/core/aggregation-pipelin... that allow for quite ad-hoc queries.

It may also mean poor schema design. You can absolutely model relational data in MongoDB without relying on joins. If you want to normalize your data don't pick Mongo.

If you aren't normalizing, how are you ensuring that you don't avoid anomalies?

Not sure if this is exactly what you're referring to, but my understanding is that picking the "right" schema for a document database to ensure that you don't end up with slower queries like mentioned elsewhere in the thread tends to benefit from thinking at a somewhat lower level of granularity than you would probably need to with a relational database. Instead of just identifying "one to one", "one to many", "many to many", it's useful to ask questions like "how 'many' is many"; as a simple example, if the "many" in "one to many" is on the order of 10, maybe it makes sense to embed them in an array rather than use a separate collection for them. It can also help to start from thinking about the types of queries you might want and then designing the schema based on them rather than starting by deciding on the schema and then having the queries be based on that; if you're going to want certain data to be accessed at the same time, you're probably going to find some way to store it together.

Proper NoSQL design has a different perspective, it asks you "well, so what if you have an anomaly?". One example could be movies, with actors, producers, genres, etc. that could all be in separate tables in a relational database, or each movie could be a document in a document database. Now let's imagine that an actor changes their name a few years after the movie is released. Is it important to go back and change the name of the actor in each of the movies? Maybe, maybe not. Certainly you can't change the credits inside the movie itself. Maybe it's sufficient to just have a link to the actor's page, where the actor's up-to-date name is. Maybe it's insufficient if the actor will get upset that their name wasn't updated across their old movies.

You choose relational models when anomalies are unacceptable and non-relational models when anomalies are acceptable.

Thanks! That is a very useful example.

That makes sense. I tend to go away from non-relational models for this reason, but it's definitely a matter of risk management.

agreed, you can have reference fields to other collections and filter by them or query related data. But I wouldn't say its designed for relational workloads. Similarly I wouldn't use MongoDB for graph queries, even though it has an operator for just that.

> agreed, you can have reference fields to other collections and filter by them or query related data.

Yes it has that, but that's an anti-pattern. What I'm referring to is schema design. In Mongo you denormalize and store relational data in the same document as references. In SQL you a record might be split up into 3 tables, in Mongo that should all live in the same document. If you can't embed and need lookups, you should avoid Mongo and other nosql DBs

No, our app didn't use JOINs for the main work. But, as I wrote, every now and then we had some one-off requests that required either JOIN or GROUP BY or both.

With postgres+JSONB we could do both at the same time on live data.