Thanks @lusewell . Unfortunately — as you suggest — I don't think there's much we can do — but this does seem like a bad bug.

It might be worth checking out numbagg — https://github.com/numbagg/numbagg — which we use for fast operations that bottleneck doesn't include. Disclaimer that it comes from @shoyer , and I've recently given it a spring cleaning. To the extent this isn't fixed in bottleneck, we could offer an option to use numbagg, though it would probably require a contribution.

If you need this working for now, you could probably write a workaround for yourself using numbagg fairly quickly; e.g.

```python In [6]: numbagg.nanmax(xarr.values) Out[6]: 0.0

or, more generally:

In [12]: xr.apply_ufunc(numbagg.nanmax, xarr, input_core_dims=(('A','B','C'),)) Out[12]: <xarray.DataArray ()> array(0.) ```