Saturday, October 29, 2011

Accessability and testability

This year's winner for the best presentation at QA&Test 2011 in Bilbao was Julian Harty and I am totally convinced that it was the right choice. Julian did not only present in a very clear style, combining theory with practical examples, but he presented a totally new view to accessability testing (@testinggeek.com).

Does your company care about accessability issues of their products?

We don't! And I don't know of anyone else. It's expensive and normally not worth the effort. Public institutions, of course, need to care about this, but private companies? Why should they care?

Julian's brilliant idea consists in using accessability technology to improve testability of our products! (see his slides, 18)

Brilliant of course because he found another use for accessability which is testability (and search engine optimisation see his slides, 18). Everybody is happy and the product gets improved in various manners: More SEO for the management, more testability for testers, better usability and accessability for the users.

In Germany we have a saying for this: it's like a kinder surprise!
(because it contains - as advertised - three things at the same time: amazement, chocolate AND a toy :-) )

Monday, October 10, 2011

Effort estimates - be brave, invent your own

Inspired by Michael Feathers' Working Effectively With Legacy Code (on google books) I felt brave enough to confront my latest "homework" in testing: "Write a detailed test plan about how to test that code!" The technical term would be "retrofitting tests for oo code" I understand.

Now, unfortunately I was rather restrained because some premises were put up that made many of Feathers' advices impossible to follow:

1.) Tests will be seperated from the developers' code, into their own projects.
2.) Only public members are relevant.
3.) You cannot break dependencies.
4.) You need to mainly write characterisation tests.
5.) My boss needs an idea about cost estimation.


So how did I start? Well, first I used that good old tranquilizer: counting. So I wrote a program using System.Reflection (.NET) to list and count the following (per project):

1.) Non-abstract classes (without enums) t (note that in System.Reflection static classes are typed sealed and abstract)
2.) Methods (and constructors) m
3.) Properties and fields (without constants) p

I then loosely mimicked a formula I found on the internet (mathematik.uni-ulm.de) for calculating the effort on the following basis:

1.) A first impression cannot count each testing state (of input parameters/the testing object), neither too much dependencies without digging too deep.
2.) There's a stable average of testing states and complexity.
3.) Objects will be used as parameters, too.
4.) My boss wants a simple answer, not a maths thesis on the subject, things need to get done.

After much trying demanding attention of a very good and patient friend I have (thank you Ewald!) I came back to the formula I started from:

testing_effort = p*m*t*TEU


(where p,m,t as above and TEU refers to "testing effort unit" which would need to be interpolated)


When I applied that formula to a recent unit testing project to interpolate TEU and after that used it to calculate a first estimation, I could be sure this wasn't the right way, as my results foresaw some gigantic ressources we just don't dispose of.
So, next step was some static code analysis targeting the question how to reduce test basis. (Remember: "We cannot test everything.")

Some sensible criteria I set up with one of our developers then were e.g.

1.) If there are two overloads, look if they really do something different, kick one out (of your calculation).
2.) If there are e.g. 20 read-only properties of type string, count 1 property of type List<string>.
3.) If two classes inherit from the same abstract class, only count where overriden / where they're different, a.s.o.

The good thing about reading the source code is that you get quite a good picture of the importance of certain classes so you can prioritize the components / units at the same time.

So then I had a final list, different countings, my formula would certainly give better results:

total_testing_effort_1 = \sum_{i=1,...,7} te_i


(where te_i is the testing effort of each assembly, te_i = p_i*m_i*t_i)


The ridiculous result was: 37h. Thirty-seven hours for seven quite complicated assemblies (e.g. there was one object that could only be created by another which itself depended on a third one).

Still I didn't panic - it's just numbers!

So, I thought: to include the dependency my assemblies have on each other, I can instead calculate the testing effort of the total numbers, thus, by mutliplication, establishing a relation between those numbers that is stronger, that is

total_testing_effort_2 = (\sum t_i)*(\sum p_i)*(\sum m_i)


resulting in another ridiculous estimate of 1039 hours.

Knowing that plan had to be written, and that I couldn't live with 37h and my boss couldn't do so with 1039h I had to find some equally acceptable number. So find some mean. It's obvious that the average isn't good enough, so I used the geometric mean of those two:

total_testing_effort = sqrt(total_testing_effort_1*total_testing_effort_2)


And there it was: a magical 119,1h approx. 25 days of 8h work.

You think this is all hocus pocus? Maybe you're right. But I've found the following to be true, too:

1.) There is an effort estimate we feel good about.
2.) My boss understands the code isn't as easy to be tested as he had thought, and will probably be more forgiving when not "everything" will have been tested.

Of course, I still hope my baby will show some good approximation over time :-)

Do you have other means of estimating costs / setting up test plans for legacy code? Found mistakes or nonsense? Please comment!