Is there some formal way(s) of quantifying potential flaws, or risk, and ensuring there’s sufficient spread of tests to cover them? Perhaps using some kind of complexity measure? Or a risk assessment of some kind?
Experience tells me I need to be extra careful around certain things - user input, code generation, anything with a publicly exposed surface, third-party libraries/services, financial data, personal information (especially of minors), batch data manipulation/migration, and so on.
But is there any accepted means of formally measuring a system and ensuring that some level of test quality exists?
Pit Mutation testing is useful. It basically tests how effective your tests are and tells you missed conditions that aren’t being tested.
For Java: https://pitest.org
Edit: corrected to the more general name instead of a specific implementation.
The most extreme examples of the problem are tests with no assertions. Fortunately these are uncommon in most code bases.
Every enterprise I’ve consulted for that had code coverage requirements was full of elaborate mock-heavy tests with a single Assert.NotNull at the end. Basically just testing that you wrote the right mocks!
That’s exactly the sort of shit tests mutation testing is designed to address. Believe me it sucks when sonar requires 90% pit test pass rate. Sometimes the tests can get extremely elaborate. Which should be a red flag for design (not necessarily bad code).
Anyway I love what pit testing does. I hate being required to do it, but it’s a good thing.
I’d never heard of mutation testing before either, and it seems really interesting. It reminds me of fuzzing, except for the code instead of the input. Maybe a little impractical for some codebases with long build times though. Still, I’ll have to give it a try for a future project. It looks like there’s several tools for mutation testing C/C++.
The most useful tests I write are generally regression tests. Every time I find a bug, I’ll replicate it in a test case, then fix the bug. I think this is just basic Test-Driven-Development practice, but it’s very useful to verify that your tests actually fail when they should. Mutation/Pit testing seems like it addresses that nicely.
We are running the above pi tests with an extra (Gradle based) build plugin so that it only runs mutations for the changed lines in that pull request. That drastically reduces runtime and still ensures that new code is covered to the mutation test level we want. Maybe something similar can be done for C or C++ projects.
I’m currently working on a C++ project that takes about 10 minutes to do a clean build (Plus another 5 minutes in CI to actually run the tests). Incremental builds are set up, and work quite well, but any header changes can easily result in a 5 minute incremental build.
As much as I’d like to try, I don’t see mutation testing being worthwhile for this project outside of maybe a few isolated modules that could be tested independently. It’s a highly interconnected codebase, and I’ve personally reviewed (or written) every test, so I already know they’re of fairly high quality, but it would be nice to be able to measure.
80%. Much beyond that and you get into a decreasing return on the investment of making the tests.
I think this is a good rule-of-thumb in general. But I think the best way to decide on the correct coverage is to go through uncovered code and make a conscious decision about it. In some classes it may be OK to have 30%, in others one wants to go all the way up to 100%. That’s why I’m against having a coverage percentage as a build/deployment gate.
“When a measure becomes a target, it ceases to be a good measure”. — Goodhart’s law
But is there any accepted means of formally measuring a system and ensuring that some level of test quality exists?
Formally? No, this is basically impossible by Rice’s Theorem. There is not even a guarantee that if you have 100% test coverage, the program is good (the tests could be flawed).
This is just a natural limitation of turing completeness. You can’t decide these properties while also having full computational power. In order to decide such things, you need a less powerful mode of computation (something not turing complete) that can be analyzed more thoroughly and with more guarantees.
That makes sense, thank you. Yes, it’s specifically “test quality” I’m looking to measure, as 100% coverage is effectively meaningless if the tests are poor.