Bug#983328: pandas: sparse test_from_coo fails with scipy 1.6.1

[Drew Parsons]

On 2021-02-22 18:50, Rebecca N. Palmer wrote:
> On 22/02/2021 12:44, Drew Parsons wrote:
>> v1.2.2 is the latest pandas release
>> Has this error in test_from_coo been fixed in that latest release?
> 
> Not obviously, but I haven't tried actually running that.
> 
> Is scipy 1.6.1 intended for bullseye?

In principle it would be nice to get the scipy update into bullseye.  I 
uploaded it to experimental first to check if anything breaks.  CI 
testing from experimental is reported at 
https://release.debian.org/britney/pseudo-excuses-experimental.html . 
numba is also affected (on arm64), and qutip on amd64.

scipy 1.6.1 is supposed to be a bug-fix release, no new features.  So 
it's a bit weird that the pandas test passed with scipy 1.6.0 but fails 
with 1.6.1.

I figure the pandas developers must have worked around some bug in scipy 
which is now fixed, but that scipy fix now breaks the pandas 
work-around. My hypothesis is that they later found a more robust way of 
working with scipy which is included in pandas 1.2 (I haven't yet tested 
it either).

Arguably it's too disruptive in the middle of soft freeze to upgrade 
pandas from 1.1.5 to 1.2.2 in unstable/bullseye. We could upload pandas 
1.2.2 to experimental though.

It's not safe to upload scipy 1.6.1 to unstable until we can fix pandas. 
We don't want to risk pandas (and scipy) being removed from bullseye.  
If we can identify a specific patch or upstream commit that fixes pandas 
1.1.5, that's probably a good outcome for bullseye.  If we can't find 
any such patch then bullseye will just have to manage without scipy 
1.6.1.

Drew