# Is it OK to use the classes under test to initialize the database for the tests?

I am working at a small in-house dev department in a non-IT organisation. Me and another junior dev are creating a mid-sized CRUD web application in ASP MVC. There is no formal test process in our department. I am trying to gently introduce some best practices including unit testing, of which I only have theoretical knowledge myself.

We started with creating a nice, clean database schema. During the requirements analysis and database modelling phases, I went through piles of real data from our domain and then entered them all per hand in the database. Then the other dev got the task of creating the Model part of the project, writing the entity and repository classes needed for Linq to SQL.

I outlined my vision for the unit tests of his classes: We already have a small but consistent set of data we can use as a test data. I suggested that he automatically recreate the database from scratch, filling the tables per SQL script, and then let the tests run against the dataset. He agreed and went to work.

Now, a few weeks later, he has for some reason decided that he doesn't like working with the data as it is. He initializes the database by re-creating the schema, and then fills it using our Linq classes and their connections. He does not pay attention to keeping the data in a state which is sensible for the domain. Example: in my data, a mouse could have a mutation expressed in the mammary glands. In his data, a drosophila can have a mutation expressed in the mamary glands, because he just set random connections between records from the proper tables, without caring for business meaning.

I see two problems with this approach.

1. I think that the tests supposed to verify the work of our Linq classes should not use the Linq classes for initialization of the data set. From my point of view, this is a bad case of circular reasoning.
2. I would like to keep one set of test data which is applicable for most types of test, maybe with a few necessary variations. While there is no reason why the data used for unit testing the database connection should be the same as the data used for tests involving the users, there is also no reason to make it different. And the users cannot work with data which tells them that a drosophila has mammary glands. So, I would like us to adapt our tests to work with the old data set.

My co-worker is unhappy, because my suggestions mean extra work load. He feels that they are completely unnecessary.

My questions for you: First, is his approach wrong (so we should change it), or is it right (maybe one of many possible right approaches)? Second, is my own suggestion right or wrong? Third, if his approach is so wrong we should change it, what arguments can I use to convince him? (He is absolutely no convinced by my "circular reasoning" argument).

Of course if you don't implicitly trust your Linq classes, you should test them.

Under which circumstances would I trust my Linq classes? Do I need to test them, or not?

He/she also said:

One way to test the Linq classes would be to write the data using Linq and then read the data using some other means

This is exactly what I was planning to do. My question was: Why? Why cannot I test the Linq classes with themselves? I thought that it is so obviously a bad idea that it doesn't need an explanation. But now that my co-worker does not believe that it is a bad idea, I realised that I have no arguments against it. So, what are the arguments for this view?

• I'm not User246, but you can trust your Linq classes when you know the data they send and return is correct. You can do this by multiple methods including using the classes to insert or update data and then manually looking in the database. Or by using an ADO connection to retrieve and check. If you have a Linq class that pulls the whole record, that's a good one to use to check a Linq update. – Kate Paulk Aug 6 '13 at 12:13

Welcome to SQA, Rumi P. It sounds like you have a bootstrap problem rather than a chicken-and-egg problem. There is nothing intrinsically wrong with using software under test to create your test data, especially if it lets you write and maintain tests more easily.

Of course if you don't implicitly trust your Linq classes, you should test them. One way to test the Linq classes would be to write the data using Linq and then read the data using some other means (e.g. hard-coded SQL), and vice versa.

(Edit)

There is an underlying testing principle here that I did not bother to articulate, so here it is: if your system under test communicates with the outside world (e.g. a database or a network connection), and you only test the system using its own operations, all you have verified is that the system is consistent with itself.

Here is an example. Let's say you want to test a class that reads and writes files. It has two operations: readFile and writeFile. readFile takes a filename and returns an array of bytes representing the file contents. writeFile takes a filename and an array of bytes and creates that file. Now, one way to test this would be to fill up an array of bytes, write to a file using writeFile, read the file back using readFile, and verify that the outgoing array is equal to the incoming array. Certainly, if that test fails, you have a problem.

Now consider this implementation of readFile and writeFile:

function writeFile(filename,byteArray):
//... write byteArray to file /home/bob/test.txt
return

byteArray = contents of file /home/bob/test.txt
return byteArray


readFile and writeFile allegedly communicate with an external system -- in this case, the file system -- but can be implemented in such a way that you cannot verify that fact using the system under test's operations. The system is consistent with itself but is not consistent with the outside world.

I hope you can infer how this relates your question (and my answer) about the Linq classes.

• Thank you for your answer. Maybe I did not write my question clearly enough, but your second paragraph is repeating what I wanted to ask, so maybe you could elaborate on why it is so? I added an update to the question body to make it more clear. Would still have upvoted, but don't have the rep yet. – Rumi P. Aug 6 '13 at 8:45
• I updated my answer to try to address your concerns. – user246 Aug 6 '13 at 19:24

Welcome to SQA.

As @user246 says, there's nothing intrinsically wrong with using your application under test to generate your test data - although you probably do want to have tests that validate the data you generate that way.

While you technically don't need to have test data reflecting real usage conditions there are times and cases where this is a good thing. Some of the factors you might want to consider are:

• If your application is going to use more or less the same data for all users, you probably want your test data to reflect real usage conditions
• If you have a large number of configuration settings that can dramatically change the behavior of any part of the application, you want your test data to reflect as many of the possible configurations as you can reasonably contain in a single data source - to use the example I've got years of familiarity with, an application that can be used in any of multiple countries has to support all the tax scenarios used in any of those countries, so the test data has at least one product of each type configured with each potential tax scenario rather than reflecting typical customer data
• If you need to validate calculations, it helps to use starting values that make the calculations easier to validate by hand when you're generating your data - in the tax example above, the test data typically used a base price of \$10 for each product or multiples of that because it's a whole lot easier to validate tax calculations when you're starting with a whole number.
• If you use the application to generate the data, it helps to have a separate validation test that does a table-crawl and checks all the non-dynamic fields (you don't want to be validating timestamps against a static baseline)

As always with this kind of thing, there isn't a "right" or "best" answer - the one that works for you and your fellow team members is fine, and can be refined as you find shortcomings with it. No matter how well you plan out your tests, you will find things you missed, so as long as you have a decent starting point, you'll be fine.

Possibly the biggest thing you want to look at is what, precisely, you're testing with the way you're doing things. As a general rule, you don't want to spend a lot of time on something that isn't your target - if you can get to your target faster and with less effort by taking shortcuts, take shortcuts (for example, if you're testing that importing data from an excel spreadsheet functions as intended, it makes more sense to use a static spreadsheet in a known location than to generate the spreadsheet in your tests - you don't need to test programmatically creating a spreadsheet, you only need to test the import).