@jhamman The problem is these lines in Dataset.__getitem__: https://github.com/xray/xray/blob/v0.6.0/xray/core/dataset.py#L615-L619

We should probably be checking for is_hashable(key) instead of converting to an ndarray, where is_hashable looks something like this: http://stackoverflow.com/a/3460725/809705

I started working on a second PR for this issue and ran across this issue:

``` Python In [1]: import numpy as np In [2]: import xray In [3]: data = np.random.random((2, 2)) In [4]: da = xray.DataArray(data) In [5]: ds = xray.Dataset() In [6]: ds[(4, 5)] = da In [7]: print(ds)

traceback shortened

KeyError Traceback (most recent call last) /Users/jhamman/Dropbox/src/xray/xray/core/dataset.py in _get_virtual_variable(variables, key) 37 """ 38 if not isinstance(key, basestring): ---> 39 raise KeyError(key) 40 41 split_key = key.split('.', 1)

KeyError: 4 ```

It looks like we have something else to fix because it is possible to create a dataset that behaves like a dictionary but cant be written to disk or printed.

variable names must be hashable (and unique)

This makes datasets behave like dictionaries, which are familiar for Python users. A good thing.

raise an error in to_netcdf in the event the variable name is invalid.

@jhamman +1 for this solution

I think I'm starting to fall in line with you here. Perhaps we should just raise an error in to_netcdf in the even the variable name is invalid.

I'm conflicted about this change. My concern is that the only strict requirement for making a dataset is that the variable names must be hashable (and unique). Yes, you certainly need strings to write netcdfs -- but that's not something everyone uses, and it's usually not much more distruptive to get the error in to_netcdf rather than when creating the dataset. I don't think there are many other places in the code where we need to verify that variable name is a string?

I think this is ready. Anyone have final comments or concerns?

Nope, the elements in new_coords_names are a subset of the keys of new_variables.

do we also need to check new_coord_names?

I think you can put the Dataset checks in _update_vars_and_coords (on the keys of new_variables). That method gets called at the lowest level by everything that adds variables.

Note, for example, that you can also create new variables by passing a dictionary into Dataset.merge or Dataset.update.

82bcd8b follows the same rules as for the Dataset key names as the DataArray.name attribute.

c = xray.Dataset({'a': 1, ('b', 'b'): 2})

now raises a TypeError.

Were you planning on also adding these checks to the keys of a Dataset in this PR? Because I believe you can still do this:

c = xray.Dataset({'a': a, ('b', 'b'): b})

and it won't raise an error.