This article seems like a very long-winded and complicated way to say that we should write tests. Am I missing something here? Wouldn't most developers write tests when creating algorithms, let alone something relating to finance as tax calculations? Yes, you should reproduce a defect by writing a failing tests first.
Where I hoped/thought this piece would go was to expand on the idea of error-prone[1] and apply it to the runtime.
I thought it was interesting - not revolutionary but updated my thinking a bit.
Writing a failing test that reproduces a bug is something I learned pretty early on.
But I never consciously thought about and approached the test as a way to debug. I thought about it more of a TDD way - first write tests, then go off and debug/code until the test is green. Also practically, let's fill the gap in coverage and make sure this thing never happens again, especially if I had to deal with it on the weekend.
What was interesting to me about this was actively approaching the test as a way of debugging, designing it to give you useful information and using the test in conjunction with debugger
This reminds me of a talk that Leslie Lamport (author of LaTeX & prominent figure in the field of distributed computing) gave recently [1]. I remember him arguing that the difficult part in writing code is not to determine what code to write to compute something, but to determine what this something is in the first place. "Logic errors" are really about valid algorithms that end up computing the wrong thing - they're gonna compile, they're gonna run, but they won't do what you want them to do.
One example he gives is computing the maximum element in a sequence of numbers. This is something trivial to implement but you need to decide what to do with the obvious edge case: empty sequences. One solution is to return some kind of error or exception, but another is to extend what we mean by the largest element in a sequence the way mathematicians typically do. Indeed, the maximum function can be extended for empty sequences by letting max([]) := -infinity, the same way empty sums are often defined as 0, and empty products as 1. The alleged benefit of following the second approach is that it should lead to simpler code/algorithms, but it also requires more upfront thinking.
Closely related are in-code assertions. I remember when I used to liberally use asserts inside a code (and you could disable them for production) to check pre-conditions, post-conditions, or any invariants. Nowadays, I don't think the pattern is recommended anymore, at least in certain popular languages.
It's not recommended as much anymore because of unit tests. Instead of peppering the code with asserts, you build tests based on those assertions. You don't have to worry about turning it off in production because the tests are separate, and you also don't have to worry about manually triggering all the various asserts in a dev build, because the test runs are doing that for you even before a build is published.
How do you determine if your tests are good at finding logic errors?
Mutation testing. Introduce artificial logic errors and see if your tests find them.
Disappointed the article didn't go into this. You can even use mutation as part of a test generator, saving the (minimized) first test input that kills a mutant. You still need some way of determining what the right answer was (killing the mutant just involves seeing it does something different from the unmutated program.)
recroad|9 months ago
Where I hoped/thought this piece would go was to expand on the idea of error-prone[1] and apply it to the runtime.
https://github.com/google/error-prone
simplesort|9 months ago
Writing a failing test that reproduces a bug is something I learned pretty early on.
But I never consciously thought about and approached the test as a way to debug. I thought about it more of a TDD way - first write tests, then go off and debug/code until the test is green. Also practically, let's fill the gap in coverage and make sure this thing never happens again, especially if I had to deal with it on the weekend.
What was interesting to me about this was actively approaching the test as a way of debugging, designing it to give you useful information and using the test in conjunction with debugger
jeremyscanvic|9 months ago
One example he gives is computing the maximum element in a sequence of numbers. This is something trivial to implement but you need to decide what to do with the obvious edge case: empty sequences. One solution is to return some kind of error or exception, but another is to extend what we mean by the largest element in a sequence the way mathematicians typically do. Indeed, the maximum function can be extended for empty sequences by letting max([]) := -infinity, the same way empty sums are often defined as 0, and empty products as 1. The alleged benefit of following the second approach is that it should lead to simpler code/algorithms, but it also requires more upfront thinking.
[1] https://www.youtube.com/watch?v=tsSDvflzJbc
pkoird|9 months ago
just6979|9 months ago
esafak|9 months ago
pfdietz|9 months ago
Mutation testing. Introduce artificial logic errors and see if your tests find them.
Disappointed the article didn't go into this. You can even use mutation as part of a test generator, saving the (minimized) first test input that kills a mutant. You still need some way of determining what the right answer was (killing the mutant just involves seeing it does something different from the unmutated program.)
codr7|9 months ago