Hello,

I have written a small wrapper around pandas internal methods https://gist.github.com/0x0L/ef78c80a42892c0f832c91357914a5a4

Missing methods are those involving list or list of arrays (join, split, partition, ...)

Let me know if there's enough interest to turn this into a full commit

With cftime installed

```python import numpy as np import pandas as pd import xarray as xr

da = xr.DataArray(np.random.randn(5000, 500)) df = da.to_pandas()

with pandas

%time df_stacked = df.stack()

Wall time: 48.3 ms

%time df_unstacked = df_stacked.unstack()

Wall time: 368 ms

with xarray

%time da_stacked = da.stack(stacked_dim=('dim_0', 'dim_1'))

Wall time: 1.03 s

%time da_unstacked = da_stacked.unstack('stacked_dim')

Wall time: 78.2 ms

```

prun points to CFTimeIndex(index) in _maybe_cast_to_cftimeindex

The behaviour is also incorrect for empty indexes: ```python da[:0].stack(dim=['dim_0', 'dim_1']).dim.to_index()

CFTimeIndex([], dtype='object', name='dim')

```

Hello,

I just noticed a performance drop in 0.10 after a conda update xarray

```python import numpy as np import pandas as pd import xarray as xr

np.random.seed(1234) ds = xr.Dataset({k: pd.DataFrame(np.random.randn(2500, 2000)) for k in range(20)}) mask = (np.random.randn(2000) > -0.2).astype(bool)

%timeit ds.sel(dim_0=slice(50, 1250), dim_1=mask) %timeit ds[0].sel(dim_0=slice(50, 1250), dim_1=mask)

xarray 0.9.6 -> 120 ± 0.4 ms, 4.2 ± 0.02 ms

xarray 0.10 -> 190 ± 0.4 ms, 6.8 ± 0.03 ms

```

This was run in a docker image. Strangely I can't reproduce it natively on macos (performance is the same as in 0.10 in docker for both versions). On a window box, with similar but "real" netcdf dataset performance is halved.

Consider

```python type(np.array([1.0]).mean())

-> numpy.float64

type(pd.Series([1.0]).mean())

-> float

type(xr.DataArray([1.0]).mean())

-> xarray.core.dataarray.DataArray

```

The issue is that this dimensionless data array won't be cast into float by numpy/pandas when constructing a new ndarray/dataframe. You'll have to do it explicitly. Not a big deal but it feels weird.

I'm sure there's a real technical reason (keeping metadata ?) behind this behavior but I couldn't find any discussion about it.

Hi,

I think xarray is missing a rank function. Is there any reason not to expose a wrapper to bottleneck.nanrankdata ?

See also https://github.com/pydata/xarray/issues/1635

[edit] Although moving rank is mentioned in the whats-new for v0.9.2 I wasn't able to find that functionality nor a trace of it in the code :)