This bug appears to have been fixed, as of Xarray 0.15.0.

Yes, I think it makes sense to couple it with keep_attrs=True. One could question whether the name should be considered separate to the attributes (and therefore have an additional keyword argument for keeping the name), but I think in most cases it makes sense to keep the name if you're already keeping the attributes, and to drop the name if you're dropping the attributes.

Now that https://github.com/Unidata/netcdf4-python/pull/778 has been merged, it should be a bit easier to support this in xarray too. Though as previously mentioned, it will require no longer calling var.set_auto_chartostring(False) for compound types.

It looks to me like #1092 is about a Dataset-like object which can contain groups and sub-groups. Here we have a simpler issue: the Dataset can still be a flat object containing a single group, but it should respect the scope of netCDF dimensions. This means that any dimensions which are mentioned but not visible in the group being written should be searched for and copied (linked?) from a parent group, up to and including the root group if the dimensions reside there.

The problem isn't as bad with a smaller example (though the runtime is doubled). I've attached a minimum working example, which seems to suggest that maybe there was a problem with xarray creating a MultiIndex and duplicating all the data? (I've left in input() to allow checking memory usage before the program exists, but there isn't much difference in this example). xrmin.py.txt

Edit by @shoyer: added code from attachment inline: ```python

!/usr/bin/env python3

import time import sys import numpy as np import xarray as xr

ds = xr.Dataset() ds['data1'] = xr.DataArray(np.arange(1000), coords={'t1': np.linspace(0, 1, 1000)}) ds['data1b'] = xr.DataArray(np.arange(1000, 2000), coords={'t1': np.linspace(0, 1, 1000)}) ds['data2'] = xr.DataArray(np.arange(2000, 5000), coords={'t2': np.linspace(0, 1, 3000)}) ds['data2b'] = xr.DataArray(np.arange(6000, 9000), coords={'t2': np.linspace(0, 1, 3000)}) if sys.argv[1] == "nodrop": now = time.time() print(ds.where(ds.data1 < 50, drop=True)) print("Took {} seconds".format(time.time() - now)) elif sys.argv[1] == "drop": ds1 = ds.drop('t2') now = time.time() print(ds1.where(ds1.data1 < 50, drop=True)) print("Took {} seconds".format(time.time() - now)) input("Press return to exit") ```

data1 and data2 represent two stages of data acquisition within one "shot" of our experiment. I'd like to be able to group each shot's data into a single dataset.

I want to extract from the dataset only the values for which my where() condition is true, and I'll only be using DataArrays which share the same dimension as the one in the condition. For example, if I do: ds_low = ds.where(ds.data1 < 0.1, drop=True) I'll only use stuff in ds_low with the same dimension as ds.data1. So in my case extracting the data with the shared dimension using ds.drop(<unused dim>) is appropriate.

It would be nice to have xarray throw a warning or error to prevent me chomping up all the RAM in my system if I do try to do this sort of thing though. Or it could simply mask off with NaN everything in the DataArrays which have a different dimension.

Give me a second to provide a minimal working example.