Code Sample, a copy-pastable example if possible

```python import xarray as xr from glob import glob

This refers to a large multi-file NetCDF dataset

file_list = sorted(glob('~/Data///*.nc'))

dataset = xr.open_mfdataset(file_list, decode_times=True, autoclose=True, decode_cf=True, cache=False, concat_dim='time')

At this point, the total RAM used by the python process is ~1.4G

Select a timeseries at a single point

This is near instantaneous and uses no additional memory

ts = dataset.sel({'lat': 10, 'lon': 10}, method='nearest')

Convert that timeseries to a pandas dataframe.

This is where the actual data reading happens, and reads the data into memory

df = ts.to_dataframe()

At this point, the total RAM used by the python process is ~10.5G

```

Problem description

Despite the fact that the resulting dataframe only has a single lat/lon point's worth of data, a huge amount of RAM is used. I can get (what appears to be) an identical pandas DataFrame by changing the final line to:

python df = (ts * 1.0).to_dataframe() which reduces the total RAM to ~2.2G (i.e. 0.6G additional RAM for that single line vs 9G additional RAM). No type conversion is taking place (i.e. ts and ts * 1.0 both have identical data types)

Expected Output

I would expect that to_dataframe() would require the same amount of memory whether or not it was multiplied by 1.0. I'm aware there could be a good reason for this, but it took me by surprise somewhat.

Output of xr.show_versions()

commit: None python: 3.5.2.final.0 python-bits: 64 OS: Linux OS-release: 4.15.0-36-generic machine: x86_64 processor: x86_64 byteorder: little LC_ALL: None LANG: en_GB.UTF-8 LOCALE: en_GB.UTF-8 xarray: 0.10.7 pandas: 0.23.1 numpy: 1.13.3 scipy: 0.17.0 netCDF4: 1.4.1 h5netcdf: None h5py: None Nio: None zarr: None bottleneck: None cyordereddict: None dask: 0.19.0 distributed: None matplotlib: 1.5.1 cartopy: None seaborn: 0.8.1 setuptools: 20.7.0 pip: 18.0 conda: None pytest: None IPython: 2.4.1 sphinx: None