[fxtentacle]

I'm surprised that they didn't mention numba.jit, which solves the same basic problem as pypy (faster numpy calculations) but in a different way that is easier to mix with existing python frameworks.

For example, TensorFlow with numba preprocessing is easy, just install both packages and it'll work. TensorFlow with pypy requires a 5 hour compile and 40 GB of temporary storage. Plus some source code fiddling inside TF, if I remember correctly.

Even as an open source project, pypy should honestly consider who they're competing with for users and funding.

[oefrha]

PyPy is targeted at general purpose workloads, NumPy support is totally an afterthought, so the basic problem it set out to solve was/is definitely not faster numpy calculations. Numba on the other hand is way more targeted at numeric workloads.

Besides, I don't see why they should mention any other project in a post announcing their departure from the Conservancy. The only surprising thing is no mention of the funding model they're moving to, other than a rather vague hint, "exploring options outside of the charitable realm".

[fxtentacle]

Based on your comment, I would guess that you never tried out numba. Of course, it can also do general python and loop optimizations. And in my experience, numba worked for every case where I couldn't get pypy to work.

And I stand by my opinion that that is something that the pypy developers should consider: is this actually usable as a solution to practical problems? Or is there something else that people use instead? If so, why? Analyzing your competition is usually a good way to learn about your own strengths and weaknesses.

[oefrha]

> Based on your comment, I would guess that you never tried out numba.

Well, you guessed wrong.

> it can also do general python and loop optimizations.

Yes, it can be used in general purpose workloads, with varying degrees of success. But its main purpose is made abundantly clear:

Accelerate Python Functions

Numba translates Python functions to optimized machine code at runtime using the industry-standard LLVM compiler library. Numba-compiled __numerical algorithms__ in Python can approach the speeds of C or FORTRAN.

Built for Scientific Computing

...

https://numba.pydata.org/

> ... Analyzing your competition is usually a good way to learn about your own strengths and weaknesses.

Except this is an announcement on their funding situation, so strengths and weaknesses are completely irrelevant, unless Numba has a particularly interesting funding model. (The funding model is government grants and corporate sponsorship, so, not particularly interesting.)

[wrmsr]

> Built for Scientific Computing

I mean the thing's called 'numba' lol.

I always liken Pypy to HotSpot in that to this day the numerical performance of the latter isn't spectacular and nobody really cares - it's built to handle the harder job of making vast tangled codebases of non-numerical application code run fast, not just tight math loops which are already handled perfectly well by other more specialized tools.

[ed_elliott_asc]

If someone announces that they are exploring options then they are told “not enough info”

If someone explores other options then announces changes they are told “should have said they were exploring options”

[oefrha]

Usually people explore other options before cutting off the current lifeline.

[nhumrich]

i don't think pypy is a numba.jit "competitor". Most people I know who use it, use it for pure python things, typically web servers, and nothing to do with machine learning or data science.