Jepsen Disputes MongoDB's Data Consistency Claims

IMO, MongoDB is really good if all your indexes and your data (after sharding) can fit in memory, and you need queries that are a simple match query against a small number of keys or a range query against a date or float type. And also your data doesn't fit into a row/column structure very easily (otherwise go SQL). The aggregation framework, joins, other features are all nice to have for occasional use, but you can't be doing those 10k times a second.

It's great for things like a realtime layer of some app that merges data with a slower and more historical layer of data running on a SQL engine or something safer. Or for services that provide realtime or recent-time analytics, while storing your historical data somewhere else (see any patterns here so far? :P ). In my case the main usage is for an advertising bid and content serving engine, which was pretty much the ideal example use-case for MongoDB mentioned in books I read years ago when first learning it.

Just to note, yes the data integrity problems are "fixed", but only if you configure your instances properly and your read and write statements properly. It's not terribly hard to do, but I don't know if I would really go recommending MongoDB for newbies. If you know how to configure it properly for your data-safety needs, and would benefit from being able to have a flexible schema in early development.... I'd still maybe suggest looking at other document DBs unless you need the read/write speed Mongo can give on simple queries (and fresh projects probably do not need that)

Tone of voice does not exist in text.

RethinkDB is no longer supported, its major caveat.

I disagree. Most of the information in a database is not something I'd want to query by (otherwise I'd be indexing every column, which I'm not), certainly not as part of the normal application.

What's the point of putting, say, every single user management field into columns in a "users" table? The regular application is never going to have to relate users by their CSRF token, or what their UI language is, or any of the other dozens of incidental details associated to a user, some visible, some implementation details.

What matters are things like the username, email, name - things the app needs to actually run relational operations on.

If you look at any real application, pretty much everyone has given up on trying to keep everything relational. That would be a massive pain in the ass and you'd end up with hundreds of columns in your users table. You'll find some form of key/value store attached to the user instead. And if you're going to do that, you might as well use a json field.

And if you do use a json field with a database engine that supports it well, like PostgreSQL, you'll find that it can be indexed if you need it anyway, and querying it is easier than joining a pile of tables implementing a KVS or two. Because yes, I might want to make a report on what users' UI language is some day, but that doesn't mean it has to be a column when Postgres is perfectly happy peeking into jsonb. And I don't need an index that will just cause unnecessary overhead during common update operations that don't use it, when I can just run reports on a DB secondary and not care about performance.

I designed an application in this manner and it has turned out exceedingly well for me. We have only had about a dozen schema changes total across the lifetime of the app. One of them involved some JSON querying to refactor a field out of JSON and into its own table (because requirements changed) and that was no problem to run just like any other database update. If we need to move things to columns we will, but starting off with an educated guess of what will need to be a column and dumping everything else into JSON has undoubtedly saved us a lot of complexity and pain.

One of our tables is just a single primary key and then a json blob. It stores event configuration. Why? Because it's always loaded entirely and there is never any reason to run relational ops on it. It's a huge json blob with lots of little details, only edited administratively, with nested fields and sub-fields (which is way more readable than columns, which are a flat namespace), including a sub-document that is actually a structure that drives generation of a pile of HTML forms. If I'd tried to normalize this I would have ended up with a dozen tables, probably more complexity than the entire rest of the DB, and a huge pile of code to drive it all, and every single application change that added an admin knob would've had to involve a dabatase schema change... All for what? Zero benefit.

Then you have to follow the principle of charity and assume they're being nice.