MongoDB vs. Cassandra

Lots of reads in every query, fewer regular writes

Both databases perform well on reads where the hot data set fits in memory. Both also emphasize join-less data models (and encourage denormalization instead), and both provide indexes on documents or rows, although MongoDB's indexes are currently more flexible.

Cassandra's storage engine provides constant-time writes no matter how big your data set grows. Writes are more problematic in MongoDB, partly because of the b-tree based storage engine, but more because of the multi-granularity locking it does.

For analytics, MongoDB provides a custom map/reduce implementation; Cassandra provides native Hadoop support, including for Hive (a SQL data warehouse built on Hadoop map/reduce) and Pig (a Hadoop-specific analysis language that many think is a better fit for map/reduce workloads than SQL). Cassandra also supports use of Spark.

Not worried about "massive" scalability

If you're looking at a single server, MongoDB is probably a better fit. For those more concerned about scaling, Cassandra's no-single-point-of-failure architecture will be easier to set up and more reliable. (MongoDB's global write lock tends to become more painful, too.) Cassandra also gives a lot more control over how your replication works, including support for multiple data centers.

More concerned about simple setup, maintenance and code

Both are trivial to set up, with reasonable out-of-the-box defaults for a single server. Cassandra is simpler to set up in a multi-server configuration since there are no special-role nodes to worry about.

If you're presently using JSON blobs, MongoDB is an insanely good match for your use case, given that it uses BSON to store the data. You'll be able to have richer and more queryable data than you would in your present database. This would be the most significant win for Mongo.

Why choose between a traditional database and a NoSQL data store? Use both! The problem with NoSQL solutions (beyond the initial learning curve) is the lack of transactions -- you do all updates to MySQL and have MySQL populate a NoSQL data store for reads -- you then benefit from each technology's strengths. This does add more complexity, but you already have the MySQL side -- just add MongoDB, Cassandra, etc to the mix.

NoSQL datastores generally scale way better than a traditional DB for the same otherwise specs -- there is a reason why Facebook, Twitter, Google, and most start-ups are using NoSQL solutions. It's not just geeks getting high on new tech.

I've used MongoDB extensively (for the past 6 months), building a hierarchical data management system, and I can vouch for both the ease of setup (install it, run it, use it!) and the speed. As long as you think about indexes carefully, it can absolutely scream along, speed-wise.

I gather that Cassandra, due to its use with large-scale projects like Twitter, has better scaling functionality, although the MongoDB team is working on parity there. I should point out that I've not used Cassandra beyond the trial-run stage, so I can't speak for the detail.

The real swinger for me, when we were assessing NoSQL databases, was the querying - Cassandra is basically just a giant key/value store, and querying is a bit fiddly (at least compared to MongoDB), so for performance you'd have to duplicate quite a lot of data as a sort of manual index. MongoDB, on the other hand, uses a "query by example" model.

For example, say you've got a Collection (MongoDB parlance for the equivalent to a RDMS table) containing Users. MongoDB stores records as Documents, which are basically binary JSON objects. e.g:

{
   FirstName: "John",
   LastName: "Smith",
   Email: "[email protected]",
   Groups: ["Admin", "User", "SuperUser"]
}

If you wanted to find all of the users called Smith who have Admin rights, you'd just create a new document (at the admin console using Javascript, or in production using the language of your choice):

{
   LastName: "Smith",
   Groups: "Admin"
}

...and then run the query. That's it. There are added operators for comparisons, RegEx filtering etc, but it's all pretty simple, and the Wiki-based documentation is pretty good.