fyrfaras@Elizafox @whitequark @imbl Alrighty, thank you for explaining it to me. ^^ Also I just read a toot thread, below, by @xgranade that mypy doesn't use “mathematical type checking” but rather “linting heuristics” which could lead to this bad reputation you've mentioned earlier I think.
Cassandra is only carbon now (@[email protected])
As I mentioned earlier, it's a bit more difficult to have good exit strategies with ruff, given that the specs around linting are much more loose. It's even harder to have a good exit strategy for ty, even though there's good specs, because there's not a great type checker to use instead¹. ___ ¹As has been pointed out to me, mypy is, for all its strengths and weaknesses, not a type checker. It doesn't follow formal mathematical type checking rules, it follows linting heuristics.
Cassandra is only carbon now@fyrfaras @Elizafox @whitequark @imbl Like, I'm not trying to be petty or sarcastic about it. Rather, it's that I think some of my earlier (and admittedly a bit toxic) hate of mypy came from deeply misunderstanding what mypy is and is trying to do. I would have expected from docs and the description, that mypy accepts well-typed programs and rejects programs that violate typing rules — but that's not it. It rejects well-typed programs that are likely incorrect, which is what a linter does.
@fyrfaras @Elizafox @whitequark @imbl In particular, I cannot predict the behavior of mypy given knowledge of Python's type system alone. I have to know about what is and is not likely a logic error in Python.
Which is fine if you want a somewhat more rigorous linter that has the benefit of being developed under the auspices of the PSF itself. But it's not what I want or would want out of a type checker.
✧✦Catherine✦✧@whitequark @fyrfaras @Elizafox @imbl Agree, absolutely.
I made the mistake of thinking of it as one for a long time, but I have been disavowed of that notion.
@xgranade @fyrfaras @Elizafox @imbl however the problem is that the mypy website has this misinformation on the landing page
@whitequark @fyrfaras @Elizafox @imbl Ooof, yeah. Were I in the mood to start useless fights, I'd be tempted to PR a change to that page, but it'd just get people hurt for no reason.
I mean, it just isn't a type checker. That's not what it does.
@whitequark I guess part of it is that I'm trying to reconcile "cool people that I think well of work on mypy" and "mypy is manifestly unfit to serve as the thing it advertises itself as." For me, that comes down to trying to understand what mypy *is* for if it's not the thing it's billed as, if there's some kind of semantic drift there.
Were I to make that PR as a frequently loud critic, I don't think it'd come across in good faith, and, well, fair enough.
@xgranade well. how to put it. I don't care how it's fixed, just that it is
@whitequark That's completely fair.
Glyph@xgranade @whitequark I'm still not really clear on why the incorrect simplifying assumptions about python's runtime that ty makes are just the practical compromises of a sound type checker but the incorrect simplifying assumptions that mypy makes put it in an entirely different category of software
@xgranade @whitequark like "too much busywork for the amount of bugs it finds" is not an experience I have had, but we write very different kinds of code so I'm definitely not arguing with that; I could see that 'ty' or 'pylance' might do far better with your code, or might even have a more rigorous approach to certain aspects of their type algebra. but all three happily give this program (and many like it) the thumbs-up:
x = 1
y = 2
locals()['y'] = 'oops'
print(x + y)
@whitequark @xgranade this is actually why I prefer mypy, specifically because (via plugins) it can accept an extension to the type system implemented with runtime hackery, and handle Zope Interface, while none of the other ones can. pydantic also has a similar thing (and dataclasses were originally supported via the 'attrs' plugin, although everybody else eventually got on board with those).
but I guess when it comes to "helpful" warnings, mypy probably does produce a lot more of those
@whitequark @xgranade in any case, I absolutely understand “the experience of using this sucks”, I just don’t get the category distinction
@whitequark @glyph Sorry, was on a StS2 run, heh. Anyway, my understanding of the difference is that sometimes, when you report a false negative in mypy (that is, a program being rejected despite being well-typed), the response is "well, don't write that program." I struggle to put into words how off the mark that is for a type checker, but it's perfectly sensible for a linter.
@xgranade @glyph a type-checker that rejects well-typed programs is by definition buggy. if it's not fixed then it is unfit for purpose
we could have a broader discussion of what "well-typed" means for Python (it is not clear to me at all) but if the above is an accurate assessment of the consensus in mypy development then it is not a(n even aspirationally correct) type-checker
@whitequark That's what I'm getting at, yeah. For a piece of software to be a type checker, there should not be a single case *ever* of a well-typed program that is rejected *by design*.
@xgranade @whitequark do you have a canonical example of such a WONTFIX?
```
def f() -> None:
return
y = f()
```
main.py:4: error: "f" does not return a value (it only ever returns None) [func-returns-value]
Found 1 error in 1 file (checked 1 source file)
@xgranade @whitequark heh. fair enough I guess, although I am tempted to troll “that’s totally a valid type check once you realize that in certain contexts None means ‘null’ but in others it means ‘void’ “
(formally speaking, -> None is already lightly gibberish as it is an unprincipled pun for NoneType; but in support of your point, ty does accept NoneType but Mypy doesn’t)
@glyph @whitequark But that's my point... Python doesn't even have a void, but mypy acts like it does!
```
from typing import Any
def f() -> None:
pass
y: Any = f()
```
Assigning a value to a variable of type `Any` should always be allowed. That should be true in any gradually typed system.
@glyph @whitequark By way of getting at what I mean about expecting there to be rules I can follow, one of those is that if `x = y` typechecks, then `f = lambda: y; x = f()` should also type check.
That doesn't hold, though:
```
from typing import Callable
f: Callable[[], None] = lambda: None
y = f() # error
x: None = None
y = x # ok
```
That means I cannot in general refactor by extracting a value into a function that returns that value, which breaks the whole idea of a type system.
@glyph @whitequark Anyway, sorry for the long reply, but I guess the short version is that I don't think any of the above behavior is consistent with a type system, but it can be consistent with a linter.
That is, knowing the types of a program in Python is not sufficient to know if that program will pass mypy. I have to know about the context in which those types are used, which breaks the abstraction entirely. It makes mypy take on a *much* higher cognitive load.
@xgranade @whitequark No apology necessary whatsoever for a long reply, I asked a complicated question!
@xgranade @whitequark Would you consider pyright's reportSelfClsParameterName warning to also be a violation of this same rule?
@xgranade @whitequark or like… this? https://docs.astral.sh/ty/reference/rules/#ambiguous-protocol-member
that looks a lot like linting
@glyph @xgranade it is normal for a typechecker to include linting rules at the "warn" level. i mean even gcc has those
what is not normal / problematic for me in this context is having those be hard-rejects. "someone dreamed up a rule out of vibes and now i have to mangle perfectly well-formed code" is not something i accept
@whitequark @glyph Part of the ergonomic difference for me is that I do run ruff (sigh) with warnings as errors. Sometimes that results in false negatives (programs rejected that are valid). In those cases, I can leave a # noqa comment indicating why. 90% of the time that's an unused import that's actually used but it doesn't detect it, though I digress.
Declaring something as Any should be the polite way to spell "fuck off, type checker" but even *that* doesn't work in mypy... hence mangling.
@whitequark @glyph That's fair, yeah. I think it's a difference in assigning false positives versus false negatives, but regardless, in either case, we're each making active decisions about how to use data that's noisy by necessity.
@whitequark @glyph I think we have different philosophies on that, yeah... not even that I disagree, I just think of it differently? I'm happy working within what a tool wants me to do, it's something my mathy brain is OK with. Just tell me the goddamned rules.
Whereas that's something mypy does not do. The rules are, to your earlier point, vibed up.
@xgranade @whitequark Would it change your opinion if mypy categorized all its "linting" stuff in a way that made it possible to disable at one go, in the way that it seems like most of the "lint" stuff in ty is classified? Rather than having specific categories and then the generalized ad-hoc "strict" category?
@whitequark @xgranade what else would make it "a type-checker" aside from its misclassification of warnings-as-errors, though? I thought I had gotten my head around this :)
@whitequark @xgranade ah okay. if such a category were added then the entire point of it would be to have a principled distinction. you can already configure whatever categories you like otherwise
@glyph @whitequark The big one: there need to be rules that allow me to predict its behavior, and those rules need to be rules that are only sensitive to type judgments, *not* out-of-band contextual information.
If I can keep the types of a program invariant, and get different behavior depending on how I express values of those types, then the system isn't just checking the types.