@martindurant - If I understand your question correctly, I think you should be able to follow a pretty standard xarray workflow:

```Python ds = xr.Dataset() ds['your_varname'] = xr.DataArray(some_dask_array, dims=['dimname0', 'dimname1', ...], coords=dict_of_preknown_coords)

repeat for each variable you want in your dataset

ds.to_zarr(some_zarr_store)

then to open

ds2 = xr.open_zarr(some_zarr_store) ```

Two things to note:

1) if you are looking for decent performance when writing to a remote store, make sure you're working off xarray@master as #1800 fixed a number of choke points in the to_zarr implementation 2) if you are pushing to GCS, some_zarr_store can be a GCSMap.

@rabernat - in trying out your branch, I've run into this error (mentioned by @mrocklin in pangeo-data/pangeo#19):

```Python-traceback ... ~/anaconda/envs/pangeo-dev/lib/python3.6/site-packages/xarray-0.10.0_79_g7b50320-py3.6.egg/xarray/backends/zarr.py in _extract_zarr_variable_encoding(variable, raise_on_invalid) 228 229 chunks = _determine_zarr_chunks(encoding.get('chunks'), variable.chunks, --> 230 variable.ndim) 231 encoding['chunks'] = chunks 232 return encoding

~/anaconda/envs/pangeo-dev/lib/python3.6/site-packages/xarray-0.10.0_79_g7b50320-py3.6.egg/xarray/backends/zarr.py in _determine_zarr_chunks(enc_chunks, var_chunks, ndim) 134 "Zarr requires uniform chunk sizes excpet for final chunk." 135 " Variable %r has incompatible chunks. Consider " --> 136 "rechunking using chunk()." % var_chunks) 137 # last chunk is allowed to be smaller 138 last_var_chunk = all_var_chunks[-1]

TypeError: not all arguments converted during string formatting ```

As far as I can tell, reworking my chunk sizes to divide evenly into the dataset dimensions has corrected the problem.

@rabernat - following @shoyer's thoughts here and in #1753, I'm not apposed to skipping the last few failing tests and live to fight strings another day.

@rabernat - It might a little but we'll sort it out. See https://github.com/rabernat/xarray/pull/3.

@rabernat - FYI: I'm playing with your branch a bit today.

@shoyer and @rabernat, can we brainstorm what a ZarrArrayWraper would need to be compatible with the new indexing API? I'm happy to implement it but could use a few pointers to get started.

```Python class ZarrArrayWraper(BackendArray): def init(self, variable_name, datastore): self.datastore = datastore self.variable_name = variable_name array = self.get_array() self.shape = array.shape self.dtype = np.dtype(array.dtype.kind + str(array.dtype.itemsize))

def get_array(self):
    self.datastore.assert_open()
    return self.datastore.ds[self.variable_name]  # returns a zarr-array

def __getitem__(self, key):
    with self.datastore.ensure_open(autoclose=True):
        data = IndexingAdapter(self.get_array())[key]  # which indexing adapter? 
        return np.array(data, dtype=self.dtype, copy=copy)

```

@rabernat - testing should be fully functional now.

@rabernat - re backends testing, #1557 is pretty close. I can wrap it up this week.