issues

I'm using xarray in ipython to do interactive/exploratory analysis on large multi-file datasets. To avoid having too many files open, I'm wrapping my file-open code in a with block. However, this means that every time I re-run the code the multi-file dataset is re-initialised, causing xarray to re-scan every input datafile to construct the Dataset.

It would be good to have some kind of 'warm start' or caching mechanism to make it easier to re-open multifile datasets without having to re-scan the input files, but equally without having to keep the dataset open which keeps all the file handles open (I've hit the OS max file limit because of this).

Not sure what API would suit this - since it while being a useful usecase it's also a bit wierd. Something like open_cached_mfdataset which closes input files after initialisation but caches the information collected and simply assumes that files don't move or change between accesses.

When plotting a time series of maps using faceting, my memory use jumps by over 3x, from about 4GB to 14GB.

Using macOS, Python 3.6, xarray 0.9.5, jupyter notebook.

Not really a bug, but I'm finding that the scipy backend is considerably faster than the netCDF backend for netCDF 3 files (using dataset: http://rda.ucar.edu/datasets/ds093.1/). Using Anaconda python with MKL. Not sure if this is always faster, but if it is perhaps xarray should default to scipy backend for netCDF 3 files?

I'm trying to select data from a collection of weather files. Xarray opens the multifile dataset perfectly, but when I try the following selection:

``` python

cfsr_new = xr.open_mfdataset('*.grb2.nc')

lon_sel = np.array(cfsr_new.lon[np.array([3, 4, 8])]) lat_sel = np.array(cfsr_new.lat[np.array([2, 3, 4])]) time_sel = cfsr_new.time[100:200]

selection = cfsr_new.sel(lon=lon_sel, lat=lat_sel, time=time_sel) selection.to_array()

```

I get:

```

TypeError Traceback (most recent call last) <ipython-input-38-3f04c6458da2> in <module>() ----> 1 selection.to_array()

/Users/<user>/anaconda/lib/python3.5/site-packages/xarray/core/dataset.py in to_array(self, dim, name) 1847 data_vars = [self.variables[k] for k in self.data_vars] 1848 broadcast_vars = broadcast_variables(*data_vars) -> 1849 data = ops.stack([b.data for b in broadcast_vars], axis=0) 1850 1851 coords = dict(self.coords)

/Users/<user>//anaconda/lib/python3.5/site-packages/xarray/core/ops.py in f(args, kwargs) 65 else: 66 module = eager_module ---> 67 return getattr(module, name)(args, kwargs) 68 else: 69 def f(data, *args, kwargs):

/Users/<user>//anaconda/lib/python3.5/site-packages/dask/array/core.py in stack(seq, axis) 1754 1755 if all(a._dtype is not None for a in seq): -> 1756 dt = reduce(np.promote_types, [a._dtype for a in seq]) 1757 else: 1758 dt = None

/Users/<user>//anaconda/lib/python3.5/site-packages/toolz/functoolz.py in call(self, args, kwargs) 217 def call(self, args, kwargs): 218 try: --> 219 return self._partial(*args, kwargs) 220 except TypeError: 221 # If there was a genuine TypeError

TypeError: invalid type promotion

```

pathlib.Path IMHO is one of the best additions to Python. Would be nice if it were possible to open files from Path without having to cast to str

I've been trying to get around some performance issues when subsetting a set of netCDF files opend with open_mfdataset. I managed to print out the generated dask graph for one variable and it doesn't seem right - it's huge, 5000 elements, and seems to have a getitem entry for every requested element for that variable.

The code that generates this select looks roughly like:

```python

paths = WEATHER_MET['latlon'].glob('_resampled.nc') dataset = xr.open_mfdataset([str(p) for p in paths]) selection = dataset.sel(time=time_sel).sel_points(method='nearest', tolerance=0.1, lon=lon, lat=lat) selection = weights ```

and the graph for one variable in the select (the irradiance value) looks like this:

mydask.pdf

WIP to use dask vindex to point based selection