Dishing the Dirt

The four presentations at CEWT #7 were on the topic of Dirty Testing Secrets. Here's my brief summary.

According to Karo Stoltzenburg, we testers have a bad case of hubris about the uniqueness and value of our work. Not to put too fine a point on it, half of what we do is pointless and in any case could be done by someone else. Testers, she says, pride themselves on questioning, communicating and facilitating communication, and finding the important bugs but really they should find a bit of time to take a long, hard look at themselves.

Questions? She's heard better from developers, subject matter experts, product owners. Testers have no monopoly on critical thinking and people in other roles have information and experience to bring to the table that testers often won't. Communication? Sure, it's common for testers to bring people together but we're also often then an extra node in the information flow network, a contributor to the cacophony of Chinese whispers rattling round any org. And bugs? Pah, who cares that there's 37 different edge cases three levels down in some barely used dialog? Great creativity on your colour-coded mind map of a zillion test cases, oh wise testing master, but which paths through the product do users actually care about?

The resulting discussion ranged wide, covering the value of connecting other people and then perhaps dropping out, the need to keep at least one eye on the big picture, exploratory testing as a core skill of testers, not blocking others from testing, the trade off between reporting all the things and not reporting something that turns out important, teams owning quality, and the outcome for the customer being paramount. (Blog post.)

There's a famous quote in marketing, attributed to John Wannamaker: "Half the money I spend on advertising is wasted; the trouble is I don't know which half."  How could we know in advance which half of the testing not to do? I asked a similar question a few years ago: testing can’t find all the bugs, so which ones shouldn’t we look for?

You can't get to know what testing is without doing it. That's the dirty secret at the heart of Sime Simic's talk. A tester new to testing starts from a very low base, with many questions to answer about the product, the environment, the company, the users, the sales proposition, their team ... and themselves. The irony is that an experienced tester on a new project, or product, or moving to a new company has many of the same questions.

Sure, the old hand will have some experience, but they also have the extra layer of complication that they don't know which pieces of that are relevant to the new role. How can a tester get credibility and respect, learn to trade cost and value, be efficient, be effective? Only by doing and reflecting, and not rushing to get there.

Again, the conversation flowed, this time passing through getting feedback to understand how others see you and your work, building relationships with developers (particularly), the need to raise interesting things to people who matter, outcomes rather than outputs in spite of pressure to produce the latter, using the flexibility that the tester role permits to go to where the value can be found, and the granularity effect on value: how high value to a company can be low to a team or an individual.

Mark Bunce's experience report described joining a project to implement crucial business logic with a mixture of rule-based and machine learning technology. When Mark came on board the system was biased in favour of the machine learning component and performance, in terms of both accuracy and speed, was poor. There were no testers.

He built up a test team and broke down the work the system was doing to see just where the machine learning was actually adding value and where rules could do better. Testers are certainly not the only people who can apply critical thinking, but they sure helped on this project where performance is now dramatically better and the proportion of logic implemented by machine learning is dramatically lower.

After Mark's talk we dove into details of the particular system and the crude but effective statistical analysis that helped to show where the holes were. We wondered about the quantity and representativeness of the test data sets used to exercise systems like this. Random selections of deidentified production data is good, but can't be the whole story; people still need to look at the system and the outputs it produces. Another interesting angle was the ethical question of getting subject matter experts to train a machine learning system that is designed to replace them. What motivation do they have to help it to do a good job?

My own talk is written up in We Don't Know?