Update setting data pointers for Cython 3

[jreback]

Need to update the following locations

• _libs/window/aggregations.pyx: bufarr.data
• _libs/reduction.pyx: chunk.data

---

https://dev.azure.com/pandas-dev/pandas/_build/results?buildId=34888&view=logs&j=33ccdd52-c922-5ef2-8209-78215e36d994&t=046ffbec-62c2-54e3-e88a-7745f4292bb6&l=457

just started failing.

cc @pandas-dev/pandas-core @pandas-dev/pandas-triage

if anyone has insights

[alimcmaster1]

Cython alpha was released today: https://pypi.org/project/Cython/#history

Locally I get below error when building the extensions

        buf = <float64_t *>arr.data
        bufarr = np.empty(win, dtype=float)
        oldbuf = <float64_t *>bufarr.data
        for i in range((win - offset), (N - offset)):
            buf = buf + 1
            bufarr.data = <char *>buf
                 ^
------------------------------------------------------------

pandas/_libs/window/aggregations.pyx:1385:18: Assignment to a read-only property

cc @scoder

[WillAyd]

My guess is this is the result of cython/cython#3571

[WillAyd]

So IIUC assigning the data pointer has been deprecated since 2013. Quick fix might be to not use Cython alpha in CI, but in any case we probably need to fix this.

Not sure there is a true replacement in the more modern API that allows for assigning the data pointer at least according to the following mailing list:

https://mail.python.org/pipermail/numpy-discussion/2013-November/068172.html

ref cython/cython#2498 and ping @rgommers as well

[scoder]

That looks like the one evil hack that smells a bit when you implement it but otherwise works perfectly, and that drops on your feet as soon as you've forgotten about it. You already CC-ed NumPy, they are the right people to ask here. CC-ing @mattip as well, who came up with the property wrapper for these fields for Cython.

I'm happy that you didn't ask for an upstream revert of the change. :)

[bashtage]

pandas/pandas/_libs/reduction.pyx

Lines 102 to 106 in 18a7e97

	arr = self.arr
	chunk = self.dummy
	dummy_buf = chunk.data
	chunk.data = arr.data
	labels = self.labels

is also very fishy. It also seems to be at the root of a very strange behavior when group by apply is used in statsmodels/statsmodels#6603 . In this example, the wrong data was ending up in the OLS regression because it looks like it has the wrong buffer somehow. This seems to be not completely safe in the case of a deferred operation.

[mattip]

Maybe rewrite that to use bufarr = PyArray_SimpleNewFromData(1, &win, NPY_FLOAT32, buf) instead of bufarr.data = ...? As for the line pointed to by @bashtage, that also needs a rewrite to be safe.

[mattip]

The Reducer class was added in a commit in response to gh-309 in 2011. Perhaps it could be rewritten to use nditer from python or the Array Iterator from C (if the former gives performance problems).

[mattip]

Of course you could just write PyArray_Bytes(chunk) = arr.data, although that would not solve any possible bugs or race conditions due to reusing chunk inappropriately

[jreback]

ok I think we need to fix the cython version again, @alimcmaster1 would you mind doing a PR for this.

We should certainly address this older code paths. IIRC @jbrockmendel was almost able to remove this entire module, but has some perf issues.