Maximilian Roos

FYI this is changing to one hour later, to 10.30AM PT. I updated the description above and #3691 updates the ics file.

It would also be great to be able to have multiple dimensions in the grouping — for example, I test a few functions across array sizes. I'd like to benchmark...

(no particularly strong view from me, I would generally avoid global options as much as possible. I'm not sure whether this is viable on `.plot` though...)

Not sure what's causing this, but confirm I can reproduce. Any ideas?

Very much agree that we shouldn't have differences in output. We're quite well-tested for `NaN`s but admittedly not for `inf`. I'll look into fixing...

Though actually numbagg does the same as bottleneck. numbagg itself tests against bottleneck, including `inf` values. ```python ...: import xarray as xr ...: import numpy as np ...: xr.set_options(use_numbagg=False, use_bottleneck=True)...

> I would opt for the first suggestion (mentioning limited support for `inf` values). For an end user as I am, "bottleneck & numbagg have similar results, but are different...

Right, this is a numerical precision issue. `bottleneck` has the same problem, if that's installed: ``` [ins] In [1]: xr.set_options(use_numbagg=False) ...: n1=9.9e+36 ...: n2=7e+36 ...: arr=[1,1,1,n1,1,n2,n1,1, 1, 1, 1, 1,...

Thanks for the issue, but we really need a more focused example to make progress here

Thanks. We could add pyarrow to the docs' conda environment to silence this, would be a welcome small PR...