How we know our code does the right thing?
The pat, Haskell-y answer is “types”, but obviously this doesn’t cover everything, or we wouldn’t bother with quickcheck and tests at all.
Obviously we write tests at the app level, but implicitly, we are also depending on all the code that we import. The particular combination of library versions Cabal has picked might be unique to your build, and an equally unique bug might lurk in that combination.
Later, I added some extra tests to robots-txt, and was very surprised when a whole swathe of unrelated tests failed. Of course, Travis had pulled in a later version of Attoparsec, and the bug was now visible.
Stackage wouldn’t have solved that problem - Attoparsec’s tests still passed, they just missed an important case. It only showed up in the context of robots-txt’s tests. Obviously I ran my app-level tests, but they were not sufficient to catch this particular bug, and it is probably unreasonable to expect app-level tests to catch lib-level problems anyway.
Upper bounds might have alleviated this slightly, but manually managing the upper bounds down to a minor version and continually updating them every time a fix comes in is tedious in the extreme, and in any case, doesn’t cover the case that a particular combination of libraries triggers the bug. This is what makes it almost impossible to test for at the library level: there are way too many ranges of library versions you’re expected to support. Combinatorial explosion sets in very quickly. However, this is eminently possible at the app level: for any given build, there’s an explicit set of versioned packages already chosen.
What I’d like to propose is a command (maybe integrated with cabal?) that would take the current set of package versions from my app, possibly as a cabal sandbox, and for each dependency, install and run its tests in a clean copy of that sandbox. A copy is necessary because while we can expect the libraries to play well together in terms of building cleanly, we can’t expect their test dependencies to be so well-behaved: no-one could cavil if you decided you wanted a particular version of hspec or quickcheck.
In the concrete case of my app, depending on robots-txt and transitively on attoparsec: if I had run robots-txt’s tests in the context of the version of attoparsec picked by cabal, I would have found the problem immediately.
This wouldn’t be a quick operation, but computer time is cheaper than human time. It also wouldn’t need to be run every time you deploy - only when you update the frozen dependencies (which we are all doing anyway, right?).
The output of this could also be fed into the anonymous build reporting that’s coming with Hackage 2 for a fairly massive set of test data to augment the already-planned build data collection.
a quick proof of concept is at https://github.com/mwotton/marvin