> You start writing assert fibonacci(15) == ... and already you’re forced to think. What does fibonacci(15) equal? If you already know, terrific—but what are you meant to do if you don’t?
Um …duh? Get out a calculator. Consult a reference, etc. Otherwise compute the result, and ensure you've done that correctly, ideally as independent of the code under test as possible. A lot of even mathematical stuff has "test vectors"; e.g., the SHA algorithms.
> Here’s how you’d do it with an expect test:
printf "%d" (fibonacci 15);
[%expect {||}]
> The %expect block starts out blank precisely because you don’t know what to expect. You let the computer figure it out for you. In our setup, you don’t just get a build failure telling you that you want 610 instead of a blank string. You get a diff showing you the exact change you’d need to make to your file to make this test pass; and with a keybinding you can “accept” that diff. The Emacs buffer you’re in will literally be overwritten in place with the new contents:
…you're kidding me. This is "fix the current state of the function — whether correct or not — as the expected output."
Yeah… no kidding that's easier.
We gloss over errors — "some things just looked incorrect" — well, but how do you know that any differently than fib(10)?
I really like this style of testing -- code that can be tested this way is also the most fun kind of code to work with and the most likely to behave predictably.
Amazing to see Jane Street uses Emacs. And property-based testing too.
> you don’t just get a build failure telling you that you want 610 instead of a blank string
So I had to scratch my head a bit because I was thinking: "Wait, the whole point is that you don't know whether what you're testing is correct or not, so how can you rely on that as input to your tests!?".
But even though I didn't understand everything they do yet I do see at least a big case where it makes lots of sense. And it happens to be a case where a lot of people see the benefits of test: before refactoring.
> What does fibonacci(15) equal? If you already know, terrific—but what are you meant to do if you don’t?
Yeah a common one is reuse a function in the same language which you believe is correct (you probably haven't proven it to be correct). Another typical one is you reuse a similar function from another language (once again, it's probably not been proven it is correct). But if two implementation differ, you know you have an issue.
> let d = create_marketdata_processor () in
> ( Do some preprocessing to define the symbol with id=1 as "APPL" )
Typo. It's AAPL, not APPL. It's correctly used as AAPL later on.
FWIW writing tests better become a joyful experience for we're going to need a lot* of these with all our AI generated code.
Recently, I have given up on writing unit tests, instead prompting an LLM to write them for me. I just sit back and keep prompting it until it gets it right. Sometimes it goes a little haywire in our Monorepo, but I don't have to accept its changes.
i think thats the best use you can get from LLM's in programming. Doing the boring simple test code that doesn't have to meet any quality requirements. Normally on a code review I ignore unit tests, written from humans oder LLM's out of this reason.
Same, and to avoid it going haywire I wrote an agents.md file with some prompts, like how to run a test for a single file and what to do before saying "I am done".
I do it the other way around: I write the specs and unleash an agent to turn my test suite from red to green.
Each of us does half the work, the other half being done by a LLM. The difference is that I specify the desired behavior, while you leave the specification up to the LLM. A little strange if you ask me!
In my experience the lack of joy or difficulty with tests is almost always that the test environment is usually different enough from the real environment that you end up needing to kind of stretch your code to fit into the test env instead of actually testing what you are interested in.
This doesn't apply to very simple functions but tests on simple functions are the least interesting/ valuable.
I realize that the example is contrived, but what is the point of writing a test of a fibonacci function if your test harness is designed to just take whatever it tells you and updates the assert to verify that what it told you is indeed what it just told you.
This assumes the code you wrote is already correct and giving the correct answer, so why bother writing tests? If, however you accept that you may have got it wrong, figure out the expected outcome through some reliable means (in this case, dig out your old TI-89), get the result and write your test to assert against a known correct value.
I wouldn't trust any tests that are written this way.
These are great to test code that has no formal specification or oracle, so you're writing a reference implementation.
First, the test fails because there's no expected output, and you get to check the existing behaviour pretty-printed. Then, if it's correct, you approve it by promoting the diff into the source code, and it becomes a regression test.
And that's why the example is good. As you first get fib(3) then see the result, verify it and then freeze and then do for fib(20). That's how software development works that you can spot errors easier than a test could.
I was inspired by the jane street post and implemented exactly this in my Scala unit testing library uTest (http://www.lihaoyi.com/post/GoldenLiteralTestinginuTest090.h...). Can confirm that auto updating golden test assertions does make working with a test suite much more joyful than struggling with each assertion by hand
There's some cool ideas about unit testing here, and I know I'm kind of missing the point, but am I the only one who finds unit tests and documentation sort of, soothing?
Of course I love solving the initial problem / building the feature etc, but I always find unit tests a calming easy going exercise. They are sometimes interesting to think about writing, but normally fairly simple. Either way, once you're testing, you're normally on the home straight with whatever it is you're developing.
mdx[1] is another variation on this, also in the Ocaml ecosystem. It’s Ocaml’s version of documentation tests as in Elixir and Rust.
But it’s not limited to that. You can write tests in markdown files independently from your documentation. Use “dune test” to run the tests and review failures with “git diff”. Accept the changes if they are correct (changed behavior) with “dune promote”. Very nice workflow.
Since when has writing tests not been a joyful experience?
I do see a lot of useless tests out in the wild. I can see writing those not bringing any joy. That is true of any useless activity. Is that what we're thinking of here?
When writing Janet, i enjoy the judge[0] style of testing because it's so interactive. I added emacs helpers, one for running existing tests and one for updating the tests. It made for a nice repl-like experience, especially nice when writing grammars[1].
24 comments
[ 3.1 ms ] story [ 51.5 ms ] threadUm …duh? Get out a calculator. Consult a reference, etc. Otherwise compute the result, and ensure you've done that correctly, ideally as independent of the code under test as possible. A lot of even mathematical stuff has "test vectors"; e.g., the SHA algorithms.
> Here’s how you’d do it with an expect test:
> The %expect block starts out blank precisely because you don’t know what to expect. You let the computer figure it out for you. In our setup, you don’t just get a build failure telling you that you want 610 instead of a blank string. You get a diff showing you the exact change you’d need to make to your file to make this test pass; and with a keybinding you can “accept” that diff. The Emacs buffer you’re in will literally be overwritten in place with the new contents:…you're kidding me. This is "fix the current state of the function — whether correct or not — as the expected output."
Yeah… no kidding that's easier.
We gloss over errors — "some things just looked incorrect" — well, but how do you know that any differently than fib(10)?
I love determinism and plain old data.
https://github.com/pointfreeco/swift-snapshot-testing
What if writing tests was a joyful experience? - https://news.ycombinator.com/item?id=34350749 - Jan 2023 (122 comments)
> you don’t just get a build failure telling you that you want 610 instead of a blank string
So I had to scratch my head a bit because I was thinking: "Wait, the whole point is that you don't know whether what you're testing is correct or not, so how can you rely on that as input to your tests!?".
But even though I didn't understand everything they do yet I do see at least a big case where it makes lots of sense. And it happens to be a case where a lot of people see the benefits of test: before refactoring.
> What does fibonacci(15) equal? If you already know, terrific—but what are you meant to do if you don’t?
Yeah a common one is reuse a function in the same language which you believe is correct (you probably haven't proven it to be correct). Another typical one is you reuse a similar function from another language (once again, it's probably not been proven it is correct). But if two implementation differ, you know you have an issue.
> let d = create_marketdata_processor () in > ( Do some preprocessing to define the symbol with id=1 as "APPL" )
Typo. It's AAPL, not APPL. It's correctly used as AAPL later on.
FWIW writing tests better become a joyful experience for we're going to need a lot* of these with all our AI generated code.
It feels ... strangely empowering.
Each of us does half the work, the other half being done by a LLM. The difference is that I specify the desired behavior, while you leave the specification up to the LLM. A little strange if you ask me!
This doesn't apply to very simple functions but tests on simple functions are the least interesting/ valuable.
This assumes the code you wrote is already correct and giving the correct answer, so why bother writing tests? If, however you accept that you may have got it wrong, figure out the expected outcome through some reliable means (in this case, dig out your old TI-89), get the result and write your test to assert against a known correct value.
I wouldn't trust any tests that are written this way.
First, the test fails because there's no expected output, and you get to check the existing behaviour pretty-printed. Then, if it's correct, you approve it by promoting the diff into the source code, and it becomes a regression test.
It catches regressions. Which is the one thing where such semi-automated testing is most useful in my eyes.
No clue though why they gave it that weird "expect" name. Basically, it's semi-automated regression testing.
There is a time and place for golden tests, but this reads more like a case for property-based testing.
Of course I love solving the initial problem / building the feature etc, but I always find unit tests a calming easy going exercise. They are sometimes interesting to think about writing, but normally fairly simple. Either way, once you're testing, you're normally on the home straight with whatever it is you're developing.
But it’s not limited to that. You can write tests in markdown files independently from your documentation. Use “dune test” to run the tests and review failures with “git diff”. Accept the changes if they are correct (changed behavior) with “dune promote”. Very nice workflow.
[1] https://github.com/realworldocaml/mdx
I do see a lot of useless tests out in the wild. I can see writing those not bringing any joy. That is true of any useless activity. Is that what we're thinking of here?
[0] https://github.com/ianthehenry/judge
[1] https://github.com/minikomi/advent-of-code/blob/d73e0b622b26...