1. The library makes an assumption on a type (in the case of Python, that's typically "there's only one type of strings", "there's only one type of streams", sometimes "there's only one type of numbers", etc.). Since the language is loosely-typed, there's no API barrier that checks this, it's all duck-typing.

2. Tons of code is written based on this assumption.

3. Assumption changes. In this example, this happened with the Python 2 => Python 3 migration, but it also happens once in a while with smaller version increments of either Python itself or Python libraries.

4. Where is the assumption used in your code? Well, if you're lucky, you're going to find out once a piece of your code throws an error because it's attempting to use a method that doesn't exist anymore. If you're not lucky, the behavior of your code has changed subtly.

In the loosely-typed world, the only way to avoid this is to have very strict boundary checks. Which means basically manually implementing subpar type checking at the borders and unit tests and/or fuzzing just to ensure that your mock-type-checking actually does its jobs.

Of course, you can often get away without doing this as long as your code is for internal use only. But if you're writing a library and if you want your users to be able to troubleshoot problems without too many difficulties, you're doing manually what the compiler is doing for you in a statically-typed language.

Metric or imperial integers? Your unit tests failed to prevent this $300 million blunder:

https://www.wired.com/2010/11/1110mars-climate-observer-repo...

But more commonly what is going to happen is that someone (two years from now) is going to change a 'person' parameter from legacy SSN to database id and some users are going to get "you don't exist" when they show up at the hospital to get medical service.

You're right though, most folks in the JS/Ruby/Python ecosystems don't do this kind of testing. It's a recurring joke:

https://www.google.com/search?q=2+unit+tests+0+integration+t...

As for frameworks in the languages you listed...

Ruby and Python: https://github.com/HypothesisWorks/hypothesis

Elixir and Erlang: https://github.com/proper-testing/proper

Node and JS: https://jsverify.github.io/

As for real world use-cases, imagine you’re writing a program that accepts timestamps as input and has to implement branching, requirements-defined business logic based off of them. When you’re writing your unit tests you can use the requirements to select timestamps that are “known good” and “known bad”, but it’s hard to explore this state space on your own.

Same thing goes for handling unexpected inputs to certain functions. You probably don’t want to check _every_ type of input for _every_ dynamic function, but it might make sense to make sure that certain “entry points” to your program fail in the expected manner when they get poorly typed input.

Sure, but you need those tests anyway to verify that your code actually works. I agree that type systems reduce the occurrences of some classes of bugs; I'm only disagreeing with the claim that they reduce the amount of tests you need to write.

I recently ported some (quite complex) code I wrote from JavaScript to Typescript. The code has about a 2:1 test to code ratio, and a fuzzer for correctness. While porting, I ended up adding a couple “useless” assert(typeof x === ...) calls to quieten the compiler, which felt useless because my code was correct. Lo and behold, the assertion tripped in my test suite - apparently I was sometimes treating a string as an object and didn’t notice. Which was a serious issue; and could end up being a security problem for some people. My fuzzer didn’t find it because it never occurred to me to add string method names in my random data generator.

Generally I find that the bugs that are easy to find with tests and the bugs that are easy to find with static types are different. You can eventually find all bugs with a sufficiently large test suite; and with enough PhDs you can apparently formally prove everything. But you get the best bang for buck with a little of each. A few tests is much better than no tests. But in the same spirit, I find no matter how big my test suite, there’s a good chance static types will improve my code.

Typescript is far from perfect, but I sleep better at night with a type checker checking my code.

I don't think we disagree much. As your experience shows, trying to replicate type checking via unit tests is almost always impractical. Which means the tradeoff of not having static typing is not that you're writing more tests, but that there's an increased chance of bugs.

And I agree that once a JS project is large enough to have a build process, the benefits of TypeScript almost always outweigh the costs. I'd even like to see JS interpreters allow and ignore TypeScript types, so you could use it without a build step.

Why would you design a function in such a way that one of the arguments can represent so many different things? The problem here has nothing to do with testing. The problem is that the function itself is poorly designed.

If anything, the difficulty of writing a test for such a function would in itself be an indication that the function needs a refactoring.