@jhamman This looks great, thank you!

@shoyer - take another look. I have basically merged our two ideas and refactored the roundtrip tests. Tests are still failing but not for py2.7, on appveyor, or py3.6 locally.

@shoyer - ready for final review.

@jhamman I think this is something to do with string/bytes. The data gets written as Unicode strings but then read back in as bytes, which is obviously not ideal.

If you want to ignore this, like the current tests, you can use assert_allclose() which has a flag for ignoring string/bytes issues (yeah, I know it's a nasty hack). In the long run we need to figure out how to solve issues like https://github.com/pydata/xarray/issues/1638. It seems like https://github.com/Unidata/netcdf-c/issues/402 is likely relevant.

Question for those who are familiar with the scipy backend. I have a few failing tests here on the scipy backend and I'm not really sure what's going on. It seems like it could be related to the mmap feature in scipy.

Thanks! I like it: simple and in accordance with netcdf4's silent overwriting. It would be cool to describe this behavior this in the documentation somewhere, and add a test that the data is correctly overwritten maybe?

@fmaussion - I've updated the append logic slightly. I'm wondering what you think? This version more aggressively overwrites existing variables (data_vars and coords).

Netcdf4 would overwrite in this situation, and I am also in favor of overwriting as this could be quite a useful usecase:

python from netCDF4 import Dataset with Dataset('test.nc', 'w', format='NETCDF4') as nc: nc.createDimension('lat', 1) nc.createVariable('lat', 'f4', ('lat',)) nc['lat'][:] = 1 with Dataset("test.nc", "a") as nc: nc['lat'][:] = 2

@fmaussion - you bring up a good point. There are two scenarios here.

1) appending to a file with existing data variables 2) appending to a file with existing coordinate variables

I'm wondering if we should disallow 1, in favor of being more explicit. So:

```Python list_of_vars_to_append = ['var1', 'var2'] ds[list_of_vars_to_append].to_netcdf(filename, mode='a')

if either var1 or var2 are in filename, raise an error?

`` as apposed to the current behavior which would silently skip all vars already infilenameand not inlist_of_vars_to_append`.

Thanks @jhamman !

With this fix existing variables will silently be ignored and won't be written, right? Maybe the expected behaviour (or specs) of the "append" option should be written somewhere in the docs and tested to avoid future regressions like we had in https://github.com/pydata/xarray/issues/1215