RE: [agile-testing] Goal-Question Metric and Test Metrics
- View Source
>Example: You are working on Open Office Speadsheet. The requirement is to"make it work exactly like excel."
What does it mean for it to work exactly like Excel--to provide exactly the
same result for a calculation for a given formula, to a given level of
precision? To provide exactly the same API (ugh!)? Exactly the same user
interface? Exactly the same performance?
>You write a little function to create random formulas, execute them in OOand Microsoft Excel, then compare the results. Because of the tight
feedback loop, and because excel can be a bit of a black box ("make it work
like excel" - would this mean to reproduce errors as well?) - it might
actually be less waste to have a computer do the inspection automatically.
You're talking about randomized high-volume automated testing. Doug Hoffman
is a big proponent of it (e.g.
ion_part-1.asp). Cem Kaner, Pat Bond, and Pat McGee did a paper on it for
STAR in 2004 (http://www.kaner.com/pdfs/highvolCSTER.pdf; you'll probably
find pages 7, 8, and 9 on Extended Random Regression Testing to be very
interesting, and page 16 describes exactly what you suggest above). This
stuff can be enormously effective, not only for comparison between two
products, but on some other unexpected levels too. It usually depends on a
reliable, high-speed oracle (although sometimes the approach produces
interesting crashes, too. (And yes, reproducing errors might be required for
"bug-for-bug compatibility", especially in commercial environments, or where
legacy systems converse with the application under test.)
>Clock cycles and electricity are cheap - human time is expensive.Absolutely. The machine part is fast, for sure. The human part includes
the programming, the evaluation of the programming, bug-fixing, and the
evaluation of the results. If and where there is a disagreement between the
application under test and the oracle, a human has to make the decision as
to whether the decision is an error in the AUT, the oracle, or the program
that compares them, AND whether that difference matters to the degree that
we're going to change the AUT--"does this mean to reproduce errors as well?"