On: Erg: a statically typed language that is Python co...

[mdaniel]

> There's no difference between a mypy-enabled codebase and a java codebase in terms of static typing (other than that mypy actually supports a generally richer type system).

that must require some advanced usage of mypy then, because my experience was that even the most simple example got the finger from mypy that would have very obviously worked in Java

    mkdir alpha
    echo 'beta = True' > alpha/models.py
    echo 'from .models import beta' > alpha/__init__.py
    echo 'import alpha; print(alpha.beta)' > doit.py
    mypy --strict doit.py
    # doit.py:1: error: Module has no attribute "beta"

And look, I'm with you that I wish any codebase with an `__init__.py` anywhere in it (thus, what I think of as a project versus some ad-hoc python scripting whatever) would type-hint the absolute hell out of things, but claiming mypy or any such linting system is as strict as a strongly typed language always boils my blood

[joshuamorton → mdaniel]

I mean I think python's module's are kind of stupid, but even as someone who literally doesn't every use python's module system, my very first thought here was "I bet strict mode is requiring explicit exports in __all__"

And indeed[0] that's the case. It makes perfect sense that the abstract machine of typed python does not allow everything that the abstract machine of untyped python does, and java would error on trying to use a private class or method just the same.

[0]: https://mypy.readthedocs.io/en/latest/command_line.html#cmdo...

[still_grokking → mdaniel]

It's imho especially ridiculous to say some linter, that needs "hints" and annotations everywhere, is comparable to a static type system when looking at modern static languages, which have usually very strong type inference.

The point of a sound static type system is that you get a guaranty that your program is type-safe even in case not everything is "hinted properly".

With the linter approach on the other hands side it won't get better than what a dynamic language offers anyway: To be (more or less) sure everything is like intended all you can do is to run your program and pray… You can't upfront rule out for sure all kinds of errors like when using a proper static language.

[joshuamorton → still_grokking]

What is the difference between a "linter" and a "static type system"? Be specific? Pytype, an alternative python "linter" provides type inference. So it's not that. The best I can come up with is toolchain integration, which is exactly what I said.

[still_grokking → joshuamorton]

TL;DR: I would say the difference is the level of confidence in the results of type-checking you'll get.

Of course an external tool could model and check against a proper type system, in theory. So that's indeed not the point.

A proper static type system comes with a formal definition and (math-like) proves of soundness (and progress). This will give you the guaranty that a well typed program is sound.

A "linter" can't give you that guaranty. All it can do is to say "LGTM"; but your program could still crash or bug-out in some other way at runtime. That's quite a big difference imho.

Linters are more like automated tests: They may tell you something is wrong, but they don't give any guaranties that things are actually right. A proper static type system OTOH gives such guaranties.

For example I would consider TypeScript being only a very funky linter—as it does not model a sound type system. My understanding is that mypy and friends are not different. All those linters produce not only way to much false positives, no, they also produce false negatives (meaning that they let unsound programs pass; something that can not happen with a formally verified type system).

As an anecdote, I got bitten by that when trying TS (and it was a very sad experience): When playing around with TS for the first time I did not know that TypeScript is barely a linter. I was really very excited about all those things you can do with proper structural types there! I was coming from Scala 2 which has only very limited support (based on JVM reflection) for this kind of types¹. But I always loved the "objects as hash-maps" concept of JS… So I've written some small REST client for one of our back-end services in TS. Everything nicely typed, no escape hatches used. Someone also told me I have to use "stric mode" which will "catch all type errors". That looked a bit odd, but OK JS was not made to be typed in the first place. Everything looked great than. As someone coming form the purely functional Scala camp (where things work when they compile :-D), and encouraged by all those people praising TS for its "powerful type system", I've put the same amount of confidence as in Scala into the typed code I've written in TS. By this I mean you don't need to double check whether things are correct if the typer says they're correct. It's enough to check the logic of your program. You don't have to try out "manually" (this includes written tests!) every line of code. Up to than everything looked good. So I've pushed things and started to do end-to-end test. That was when I fell out of my dress. My code produced runtime errors which where actually "impossible"! The TS "type-checker" said things are such and such but at runtime they were different (which I found out only after some deep debugging sessions as this was something I did not expect in any way)! I think I've wasted one and a half day debugging this shit only to find out that TS "types" are actually worthless. If I have to try out any and every line of code manual anyway what's the point of TS? I could have written this thing in pure JS (where I have to also check any and every line manually), but it would have been much easier without "wrestling" with that quite useless "type system"… :-(

The point is: A dynamic language remains a dynamic language even if you use some linter! Without true guaranties you can't have much additional confidence in your code without testing just everything—even the parts that were already "tested" by our "type system". Actually it's even worse than without a linter if you don't know about its shortcomings and make the mistake to blindly trust such a tool; like you can do in case of a sound type system (modulo the very seldom cases of compiler bugs).

I hope I could make it now more understandable what I mean by "proper static type system".

___

¹ Things improved in Scala 3 though.

[joshuamorton → still_grokking]

> My understanding is that mypy and friends are not different. All those linters produce not only way to much false positives, no, they also produce false negatives (meaning that they let unsound programs pass; something that can not happen with a formally verified type system).

I'm pretty sure the mypy and typescript devs would disagree here. The typescript and mypy abstract machines aren't any less sound than Java or c++.

All four are abstract machines that make fundamentally similar guarantees. Cpp and Java can still encounter classcast exceptions and segfaults.

And, like, you still need tests in functional languages. Specifically, I don't think I've encountered anything like

> My code produced runtime errors which where actually "impossible"! The TS "type-checker" said things are such and such but at runtime they were different

Without either explicitly subverting the type system, or dealing with some kind of untyped deserialization along an api boundary, which you absolutely need to test in any language.

[mdaniel → still_grokking]

Did you reply to the right person? I'm with you, friend :-)

[still_grokking → mdaniel]

I think I did not object anything you've said. :-)

I only wanted to point out once more that a proper static type system is of more use than some bolted on linter. (And add the point about type inference).