1. Abstract
2. Refactoring

Our task at hand is to get as much coverage as possible when it comes to errors in the implementation. The ideal coverage is one where all dates are tested. The problem is that the number of dates is infinite. This is of course always a an issue when testing software but it may be more obvious here. If we test all dates there will we a great deal of dates that we know are correct. It should suffice to test only the dates that are incorrect. However, this number is also infinite. We will have to limit our coverage but still get as much breadth as possible. Basically we need to get back to the definition.

According to the ISO standard, in the period 4 January - 28 December the week number is always the same as the Gregorian week and the same also apply for all Thursdays. Thus, the mentioned dates can be safely excluded in our test coverage. Though we still have an infinite number of dates to cover, we are getting there. Finally, we can limit our test coverage to a set of years that can be reasonable argued to appear in the application which we are implementing.

To simplify it the resulting dates are non Thursdays 1-3 January and 29-31 December every year from 2005 - 2015.

Using the calendar in Outlook I compile a matrix of the week numbers in the year interval. The dates that are grayed are Thursdays which should not be part of the test.

	01-jan	02-jan	03-jan	29-dec	30-dec	31-dec
2005	53	53	1	52	52	52
2006	52	1	1	52	52	52
2007	1	1	1	52	52	1
2008	1	1	1	1	1	1
2009	1	1	1	53	53	53
2010	53	53	53	52	52	52
2011	52	52	1	52	52	52
2012	52	1	1	52	52	53
2013	1	1	1	52	1	1
2014	1	1	1	1	1	1
2015	1	1	1	53	53	53

The matrix can be thought of as a set of { date, week number } pairs which can easily be implemented in code.

I create a test which iterates over the set and makes assertions that the correct week was generated by the calendar. It turns out however, that only eleven of the dates generate a wrong week number. I refactor the code to only include the weeks that are wrong. I get a red bar and I see another opportunity for refactoring, the previous two tests can be integrated with this test by just adding the date - week pair to the set. The following is the resulting test:

 

Now that we have a test that fails we can come up with a solution to get the test to pass.

The implementation of weeks according to ISO 8601 (used e.g. in Sweden) is faulty in .NET Framework. In particular the GregorianCalendar and CultureInfo classes are faulty when it comes to ISO 8601.

So, how do we know that the implementation of ISO 8601 is faulty using TDD? We come up with a test that fails. By the ISO definition the first week is the week with the year's first Thursday in it. So, we know that e.g. 2003-12-31 is actually week number 01 of 2004 since it is a Wednesday and thus belongs to the first week. The following is a first attempt at a test:

 

Now we get a red bar and we know that the implementation is flawed. However, this is only a single test. If we get a green bar, is the implementation correct? No, of cource not. We need to write more tests to be sure.