You could do it by creating a flatten view of the underlying numpy array, building the multiindex manually and creating a new DataArray from the view and the multiindex.

@shoyer it should be all good now

@spencerkclark Are you ok with keeping the len > 0 test and then just moving up the assert at the top of the constructor of CFTimeIndex so that we raise before allocating the memory with result._data = np.array(data, dtype='O')

Hello @maxdow

I'm not sure I understand your issue correctly. If all you want is to keep only the points time, lat, lon for which t and/or dd is not NA you can simply use ```python z = d.stack(index=d.dims).dropna('index', how='all')

<xarray.Dataset> Dimensions: (index: 346) Coordinates: * index (index) MultiIndex - time (index) datetime64[ns] 2019-03-09 2019-03-09 ... 2019-03-09 - lat (index) float64 43.5 43.5 43.5 43.75 43.75 ... 45.0 45.0 45.5 45.5 - lon (index) float64 1.25 1.5 4.25 1.75 2.25 ... 1.75 2.0 4.0 1.5 2.0 Data variables: t (index) float64 0.6184 0.2966 1.426 0.8666 ... 0.09348 1.271 1.353 dd (index) float64 2.075 1.174 0.906 0.6451 ... 0.3491 0.1789 0.8922 You still have access to the coordinates of the valid points through `z.time`, `z.lat` and `z.lon`:python z.lat

<xarray.DataArray 'lat' (index: 346)> array([43.5 , 43.5 , 43.5 , ..., 51.25, 51.25, 51.25]) Coordinates: * index (index) MultiIndex - time (index) datetime64[ns] 2019-03-09 2019-03-09 ... 2019-03-09 - lat (index) float64 43.5 43.5 43.5 43.75 43.75 ... 45.0 45.0 45.5 45.5 - lon (index) float64 1.25 1.5 4.25 1.75 2.25 ... 1.75 2.0 4.0 1.5 2.0 ```

One can still unstack this and achieve the equivalent of multi-dimensional dropna. However in your example all the coordinates for time look equal (they shouldn't be) and this will prevent unstacking.

I hope this helps

Hello @ray306

stack/unstack in xarray is a bit different from pandas mechanics.

• From the stack docstring ``` Stack any number of existing dimensions into a single new dimension.

New dimensions will be added at the end, and the corresponding coordinate variables will be combined into a MultiIndex. `` Breaking it down,da.stack({'index': ['variable']})fails because it's trying to: 1. create a new multi-index with a single level inside (variable) [so far so good but probably not what you had in mind] 2. add a new dimension named 'index' with coordinates given by the newly created index. This part yields the error you saw sincedaalready has an index namedindex`.

To mesh the variable dimension into the index dimension, you first need to unstack it: python da.unstack('index').stack(index=('first', 'second', 'variable'))

• From the unstack docstring ``` Unstack existing dimensions corresponding to MultiIndexes into multiple new dimensions.

Parameters

dim : str or sequence of str, optional Dimension(s) over which to unstack. By default unstacks all MultiIndexes. `` Yourdaarray only has two dimensions:variableandindex. That's why you can'tda.unstack('first').first` is a coordinate (along the dimension 'index') not a dimension.

You could use da.unstack('index').stack(index=['second']) but this is somehow the same thing as da.unstack() since a multi-index with a single level is totally equivalent to the level itself.

I hope this helps you enjoy xarray and use it more.

@shoyer The api reference does not generate doc for the accessor methods. It's also missing for .dt Do you know how to do this ?

@jhamman Thank you

@shoyer Thanks for the help for the writing, most of these are your own words ^^

Should i move the implementation on Variable and use temp datasets like quantile does ?

@shoyer when the ranking dimension is missing from an array in the dataset, should we do nothing, or put 1s everywhere ? None of these looks appealing or natural.

@shoyer all good

Dropped support for Dataset. I don't think there's much use for it anyway.

@shoyer Thanks for the comment and the tips! I'll make the changes asap, hopefully this week-end

I dug it a bit further. Contrary to what I said numpy is fine. If x is a DataArray

```python np.array(x.mean())

-> array(2.0)

np.array([x.mean()])

-> array([2.0])

np.array(x)

-> ok

```

pandas behavior is more surprising

```python pd.Series(2)

ok

pd.Series(x.mean())

-> TypeError: len() of unsized object

pd.Series([x.mean()])

-> dtype=object

however for 1D x

pd.Series(x)

dtype is fine

```

Something looks a bit odd to me. But it looks more like a pandas issue

@fmaussion

In a nutshell,

python x = xr.DataArray([1,2,3]) pd.Series({'mean': x.mean()}) does not work as expected. You'll want one of

python pd.Series({'mean': float(x.mean())}) pd.Series({'mean': x.mean()}, dtype=float) pd.Series({'mean': x.mean()}).astype(float)

A few points:

• nanrankdata only support numeric types, how about wrapping rankdata with a keyword option ?
• There's also a push push method similar to pandas ffill which would be nice.
• move_rank outputs are not normalized as rankdata, this might be a bit confusing

Thanks for your answer. I'll try to find a good spot in the doc to write a line or two about this.