I had previously written why mutation analysis is not yet ready for prime time. Is there any way we can use mutation analysis now to improve our test suites?
It seems there is indeed a way to do that so long as you are willing to follow certain best practices.
Use mutation analysis in the small
When writing unit tests, only create mutants in the specific unit you are testing. That is, if you are testing a function f(), which in turn calls g(), only create mutants in function f(), and verify the following in those mutants.
Ensure new mutants are killed for each new tests added
The first idea is that, while mutation score seems to be a measure with specific range (0..100), the twin problems of equivalent and redundant mutants ensure that we can not trust the mutation score. Rather, the only thing we can trust is the absolute mutation detection numbers. The idea is that when you add a new test, make sure that it kills at least some new mutants.
Ensure that your assertions actually kill new mutants compared to test cases without assertions
When adding a new test case, make sure that your assertions are actually working, by verifying that they are able to kill more mutants when compared to running a test case without assertion.
Because you are evaluating mutants in much smaller units, there is a better chance you will be able to eliminate equivalent mutants as they occur.
Use statement deletion operator exclusively to obtain mutation score
Use statement deletion (Offutt 2013) mutation as the primary means to measure quality of test suite. The interesting thing about statement deletion is that if a statement (or an expression) can be deleted to create a semantic clone, then we can argue that the clone is better compared to the actual program because it is more concise. (If the deleted statement (or expression) is an optimization, and the semantic clone results in a worse program, it suggests that the tests should have included a performance metric for distinguishing the optimization).
Thus, if one is using deletion mutation exclusively, then 100% mutation score does mean something, and is achievable for a perfect program. If you are unable to obtain 100% with deletion mutation operators, it suggests that either your program is not perfect, or your test suite is not adequate.
The other important benefit of statement deletion mutation is that there is a consistent and standard definition (unfortunately, expression deletion operators are inconsistent in this regard). That is, you can expect tools that implement statement or expression deletion mutation to provide similar scores (although not across translation stages).