Hi, Pyrefly developer here. It's been almost a year since our alpha has been released and we've been working hard towards a stable release. We're excited to announce general availability today!
What type of deltas? We compare "conformance" on [this](https://pyrefly.org/blog/typing-conformance-comparison/) blog, which was a benchmark created by the creator of Pyright. However, it is quite specific to un-specified sections of PEP484 (things that checkers may not need to agree on but should). In terms of performance, you can see results [here](https://python-type-checking.com/typecheck_benchmark/) where we are quite a bit faster than pyright (almost 10x avg) and use far less memory (almost half).
I'm glad it's 1.0, but for me the most exciting thing is the tensor shape typing they mention in the blogpost. I've been looking for a good solution to this for ages, and the Pyrefly implementation seems to have everything I'd like - incl. static typing (unlike jaxtyping's runtime-only setup), and shape inference (so passing a 2,3 tensor through a 3,10 linear layer gives you a 2,10 output tensor).
Congrats on the release, and thanks for this nice tool !
I migrated a small python/django backend project at work from basedpyright to pyrefly 1.0 yesterday to try it, and the migration was painless: just running pyrefly init converted the basedpyright config to a pyrefly config in the pyproject.toml.
The typecheking phase went from ~3/4s to 0.5s which is very nice when ran with pre-commit hooks. It also caught some unnecessary type conversion that basedpyright missed (ex: calling str(value) on a value that is already a string).
I haven't toyed too much with the LSP in zed, but I did not feel anything was missing for now.
I will try to migrate larger backends in the following weeks to see how it goes.
Anyway, it's very cool to see a new stable and fast typechecker for Python ! :)
See the official blog post for more details: https://pyrefly.org/blog/v1.0/