Properly configured MySQL can safely be partitioned. This is what EngineYard does and what every single virtual host provider does. I fail to see a difference.

If you don't trust a service provider, you shouldn't use them. It is up to the company providing the service to instill trust in their services.

You use this word "safely" as if you had some ability to know that. You don't. Consider rewording.

The vast majority of MySQL deployments conceal the MySQL connection handler from end-users --- in the vast majority of deployments, it's the web app stack talking to MySQL, not the client. Virtually all security testing of MySQL is done to that service model.

"Partitioned" MySQL is a totally different service model, where the attacks aren't simply string escapes or column truncation through HTTP post-args, but arbitrary SQL syntax, arbitrary MySQL commands, and the MySQL network protocol itself. That service model has received very little security testing. Some data points for context:

* The second model implicates vastly more MySQL code, which has to be bug-free for this to be safe.

* No major database has survived intensive post-auth security testing; Litchfield's results on Oracle --- a database that got millions of dollars of security attention internally --- were all post-auth; that testing happened after he (presumably) beat the crap out of Microsoft SQL Server.

* Even if you "partitioned" MySQL into multiple running instances of mysqld running under different UIDs in different filesystem paths --- which is not what this article is implying --- a vulnerability in that MySQL code still gives an attacker code execution on your database server, which would be a game-over problem in a commercial app.