Unit-Testing: Database set-up for tests

Here's the general approach I try to use. I conceive of tests at about three or four levels:: unit-tests, interaction tests, integration tests, acceptance tests.

At the unit test level, it's just code. Any database interaction is mocked out, either manually or using one of the popular frameworks, so loading data is not an issue. They run quick, and make sure the objects work as expected. This allows for very quick write-test/write code/run test cycles. The mock objects serve up the data that is needed by each test.

Interaction tests test the interactions of non-trivial class interactions. Again, no database required, it's mocked out.

Now at the integration level, I'm testing integration of components, and that's where real databases, queues, services, yada yada, get thrown in. If I can, I'll use one of the popular in-memory databases, so initialization is not an issue. It always starts off empty, and I use utility classes to scrub the database and load exactly the data I want before each test, so that there's no coupling between the tests.

The problem I've hit using in-memory databases is that they often don't support all the features I need. For example, perhaps I require an outer join, and the in-memory DB doesn't support that. In that case, I'll typically test against a local conventional database such as MySQL, again, scrubbing it before each test. Since the app is deployed to production in a separate environment, that data is untouched by the testing cycle.

The best way I've found to handle this is to use a static test database with known data, and use transactions to ensure that your tests don't change anything.

In your test setup you would start a transaction, and in your test cleanup, you would roll the transaction back. This lets you modify data in your tests but also makes sure everything gets restored to its original state when the test completes.

I know you're using C# but in the Java World there's the Spring framework. It allows you to run database minipulations in a transaction and after this transaction, you roll this one back. This means that you operate against a real database without touching the state after the test finishes. Perhaps this could be a hint to further investigation in C#.