tracker: use faster type checker (pyrefly / ty / zuban) in CI

[MarcoGorelli]

Polars currently uses mypy for type-checking, but it's a bit slow on a codebase of this size (~1 minute)

Pyrefly takes about 1-2 seconds on my laptop, and it is a bit stricter on some things. ty is similar.

Before being able to introduce either of them, quite a few type issues which mypy doesn't flag need addressing. I'm looking into this. There's also a few issues which need addressing upstream - I'll report things as and when I find them

So, here's a tracking issue that I'll reference in such PRs

---

Both pyrefly and ty are quite new, so I'm not sure I'd suggest completely removing mypy for the time being. Personally I don't see the harm in running multiple type checkers in CI, especially if some are quite fast

---

pyrefly issues uncovered as a result of this:

• Object of class _lru_cache_wrapper has no attribute get when using lru-cached property facebook/pyrefly#2979
• unsafe-overlap reported for checking if Iterable is Generator facebook/pyrefly#2978
• Union of Iterator ignored when reporting invalid-yield facebook/pyrefly#2976
• false positive bad-argument-type when using metaclass facebook/pyrefly#2968
• False positive [bad-argument-type] for list(null_values.items()) facebook/pyrefly#2912
• Argument list[str] is not assignable to parameter iterable with type Iterable[LiteralString] in function map.__new__ facebook/pyrefly#3030
• Regression when using type alias on Enum facebook/pyrefly#3042 (not a blocker, have worked around it)
• Any declaration before loop ignored facebook/pyrefly#3046
• bad-specialization falsely reported when matching typevar facebook/pyrefly#3047
• __reversed__ signature ignored when using reversed and enumerate? facebook/pyrefly#3048
• * is not supported between list[dict[str, Unknown | None]] and Literal[1000] [unsupported-operation] facebook/pyrefly#3074

---

precursor PRs:

• chore(python): Improve internal typing ahead of using ty / pyrefly #27050
• chore(python): Make internal typing more precise (part ii) #27117
• fix(python): Named aggregation __structify was being ignored #27148
• chore(python): Fix test_dtype_concat_3735 not actually iterating through numeric dtypes #27178
• chore(python): Remove dead code in test_scan_lines #27213
• chore(python): Use more precise internal typing (pt. iii) #27232
• chore(python): Solve type: ignore in _AioDataFrameResult #27311
• chore(python): Resolve bad instantiations in test_iceberg #27314
• chore(python): Resolve type-ignores in udfs.py #27341
• fix(python): Remove invalid type annotation Sequence[int] from DataFrame.__setitem__ key #27355
• chore(python): Remove unused type: ignore statements #27360

---

relevant issues:

• TYP: DataFrame.__init__ accepts any input if not all optional dependencies are installed #27347

[Kevin-Patyk]

@MarcoGorelli This article might be helpful for you, especially the "Common migration issues" section. Had to do this at work and it helped a lot.

[mcrumiller]

Word of warning that OpenAI just acquired Astral, not sure what that means down the road.

[MarcoGorelli]

true, though i'd say that steps taken towards stricter typing will be worthwhile whichever next-gen python type-checker Polars ends up using

[asukaminato0721]

I also suggest give this a try, https://github.com/zubanls/zuban :)

[MarcoGorelli]

Oops didn't mean to close this with the linked pr