> While NoSQL databases may not force you into a schema for your database, you will pay a toll for this mutability in your application layer.

A while ago I read a really good point on this: There's always a schema. NoSQL / "schemaless" databases don't get rid of the schema, they just move it out of the database layer. It's still there, whether in the application or the documentation or just in your head.

In a relational database, you have one schema and you have to convert/update your data to match that schema whenever it changes.

I've done one big project many years ago that used NoSQL storage of heterogeneous "document" types -- a perfectly reasonable use-case for the technology -- and I wouldn't do it again. Although at the time it seemed crazy to create different tables to hold this data it would have made that code so much simpler now.

Mutable data types, with mutable data, and inconsistent error handling.

Burned me out in a matter of months.

> If your business is not experiencing massive growth that would require more servers and you’re only working with data that’s consistent, then there may be no reason to use a system designed to support a variety of data types and high traffic volume.

This whole article seems heavily NoSQL biased. I don't know what qualifies as high traffic volume, but Postgres has been known to handle 500,000 transactions per second, and there is stuff in the pipeline that is only going to make that faster.

http://akorotkov.github.io/blog/2016/05/09/scalability-towar...

There are very good ORMs and frameworks that make working with SQL simpler. With baked in type checking and migrations, there is not a lot of need for NoSQL. Plus PostgreSQL and it's JSON data type basically made a lot of the popular NoSQL options obsolete.

The other thing this article totally fails to address: Why NoSQL was so problematic early on. How Mongo by default would let errors fail silent in Java and other drivers.

She doesn't address the huge limitation in many NoSQL solutions: consistency. She briefly mentions ACID, but neglects the primary use case for document based NoSQL: if your data doesn't need to be current on ever read. So a blog is a great example, or any type of social network really. It really doesn't matter if you're reading a version of the comment thread from 20 seconds ago or 5 seconds ago (as long as it's not like 2 minutes ago).

The author also classifies all NoSQL as document stores, when they're not. Mongo uses a document model/concept, but others may use a model that's more like a distributed memcache or key/value store. There are things like Apache Geode which uses a distributed object store with client based schema versioning. The data model space is pretty big and they're pretty different fundamentally.

The author mentions how a lot of solutions use both, but in other posts I've looked at and other companies I've worked for, I've found that you often have to require both. In a lot of implementations, people use both a NoSQL and a relational store; one more as a fast read cache and the other for long term persistence. Then you get into all the issues with insuring consistency between the two and it's a miss.

Overall, this is a pretty basic article and not one I'd recommend. It can lead to the wrong decision for a project.

Looking at the author bio at the bottom, I believe it is "her".

Adds nothing to the conversation and is actually incorrect, as male is the default gender for pronouns.

Your reasoning seems too black and white, most the of the time, you can't know your data model ahead of time, that's why prototypes are so useful, so just using your language data structures or some less rigid datastore lets you get quicker to your data model.

Flyway is a fantastic tool for making changes to relational schemas. Your code is versioned, and you can just keep the schema migration scripts right next to it. If you need to upgrade an old schema, it just walks through the migration scripts until it's up-to-date.

If there is some portion of your model that really, truly, honestly consists of unstructured data, you can just put that in a JSON column.

You really seem to favor initial velocity and rapid iteration. Fair, I can totally understand why.

I favor planning, stability, and limiting thrash due to bad/uninformed decision making.

In my experience the rush to market only pans out when you plan on having a large influx of funding to rebuild your stack. Otherwise you are stuck holding a mess of code and trying to scale ideas that never should have gone to production in the first place.

There is a whole other side to the industry that demands provably correct solutions, and will not tolerate the "hacker culture". Not because of any petty reasons like work hours or perks, but because some applications require a level of accuracy that your average growth hacker can't meet.

http://sqitch.org/