Code Sample

```python import xarray as xr from dask.array import ones import dask from dask.diagnostics import ProgressBar ProgressBar().register()

Define a mock DataSet

dset = {} for i in range(5): name = 'var'+str(i) data = i*ones((8,79,200,401),dtype='f4',chunks=(1,1,200,401)) var = xr.DataArray(data=data, dims=('time','level','lat','lon'), name=name) dset[name] = var dset = xr.Dataset(dset)

Single thread to facilitate debugging.

(may require dask < 0.18)

with dask.set_options(get=dask.get):

# This works fine. print ("Testing immediate netCDF4 writing") dset.to_netcdf("test1.nc")

# This can be twice as slow as the version above. # Can be even slower (like 10x slower) on a shared filesystem. print ("Testing delayed netCDF4 writing") dset.to_netcdf("test2.nc",compute=False).compute()

```

Problem description

Using the delayed version of to_netcdf can cause a slowdown in writing the file. Running through cProfile, I see _open_netcdf4_group is called many times, suggesting the file is opened and closed for each chunk written. In my scripts (which dump to an NFS filesystem), writes can take 10 times longer than they should.

Is there a reason for the repeated open/close cycles (e.g. #1198?), or can this behaviour be fixed so the file stays open for the duration of the compute() call?

Output of xr.show_versions()