BUG: set_variables in backends.commons loads target dataset

[jhamman]

Problem description

In #1609 we (I) implemented a fix for appending to datasets with existing variables. In doing so, it looks like I added a regression wherein the variables property on the AbstractWritableDataStore is repeatedly queried. This property calls .load() on the underlying dataset.

This was discovered while diagnosing some problems with the zarr backend (#1770, pangeo-data/pangeo#48 (comment)).

I have a potential fix for this that I will post once the tests pass.

cc @rabernat, @mrocklin

Output of xr.show_versions()

INSTALLED VERSIONS ------------------ commit: 20f957d python: 3.6.0.final.0 python-bits: 64 OS: Darwin OS-release: 17.3.0 machine: x86_64 processor: i386 byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: en_US.UTF-8

xarray: 0.10.0+dev18.g4a9c1e3
pandas: 0.21.0
numpy: 1.13.3
scipy: 0.19.1
netCDF4: 1.3.0
h5netcdf: 0.5.0
Nio: None
zarr: 2.1.4
bottleneck: 1.2.1
cyordereddict: None
dask: 0.15.4
distributed: 1.19.3
matplotlib: 2.0.2
cartopy: 0.15.1
seaborn: 0.8.1
setuptools: 33.1.0.post20170122
pip: 9.0.1
conda: None
pytest: 3.2.3
IPython: 5.2.2
sphinx: 1.6.3

[shoyer]

Oh my... this is a good reason why we should get ride of properties that do non-trivial computation.