attrs empty for open_mfdataset vs population for open_dataset

[pwolfram]

Previously, a dataset would store attrs corresponding to netCDF global attributes. For some reason, this behavior does not appear to be supported anymore. Using this dataset: https://github.com/pydata/xarray-data/raw/master/rasm.nc

In [1]: import xarray as xr

In [2]: ds = xr.open_dataset('rasm.nc')
/Users/pwolfram/src/xarray/xarray/conventions.py:386: RuntimeWarning: Unable to decode time axis into full numpy.datetime64 objects, continuing using dummy netCDF4.datetime objects instead, reason: dates out of range
  result = decode_cf_datetime(example_value, units, calendar)

In [3]: ds
Out[3]: 
<xarray.Dataset>
Dimensions:  (time: 36, x: 275, y: 205)
Coordinates:
  * time     (time) object 1980-09-16T12:00:00 1980-10-17 ...
    yc       (y, x) float64 16.53 16.78 17.02 17.27 17.51 17.76 18.0 18.25 ...
    xc       (y, x) float64 189.2 189.4 189.6 189.7 189.9 190.1 190.2 190.4 ...
  * y        (y) int64 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ...
  * x        (x) int64 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ...
Data variables:
    Tair     (time, y, x) float64 nan nan nan nan nan nan nan nan nan nan ...
Attributes:
    title: /workspace/jhamman/processed/R1002RBRxaaa01a/lnd/temp/R1002RBRxaaa01a.vic.ha.1979-09-01.nc
    institution: U.W.
    source: RACM R1002RBRxaaa01a
    output_frequency: daily
    output_mode: averaged
    convention: CF-1.4
    references: Based on the initial model of Liang et al., 1994, JGR, 99, 14,415- 14,429.
    comment: Output from the Variable Infiltration Capacity (VIC) model.
    nco_openmp_thread_number: 1
    NCO: 4.3.7
    history: history deleted for brevity

In [4]: ds = xr.open_mfdataset('rasm.nc')

In [5]: ds
Out[5]: 
<xarray.Dataset>
Dimensions:  (time: 36, x: 275, y: 205)
Coordinates:
  * time     (time) object 1980-09-16T12:00:00 1980-10-17 ...
    yc       (y, x) float64 16.53 16.78 17.02 17.27 17.51 17.76 18.0 18.25 ...
    xc       (y, x) float64 189.2 189.4 189.6 189.7 189.9 190.1 190.2 190.4 ...
  * y        (y) int64 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ...
  * x        (x) int64 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ...
Data variables:
    Tair     (time, y, x) float64 nan nan nan nan nan nan nan nan nan nan ...

The attributes for open_mfdataset are missing whereas in previous versions of xarray I do not believe that this was the case because one of my scripts is failing because it does not obtain attributes when using the open_mfdataset initialization.

@shoyer and @jhamman, is this the expected behavior and was the prior behavior simply an unspecified side-effect of the code vs a design decision? My preference would be to keep as many attributes as possible when using open_mfdataset to best provenance the dataset, i.e., ds.attrs should not be empty following initialization.

[shoyer]

This was certainly not an intentional change, but on the other hand, the original behavior was not tested or intentionally specified, either.

It would be good to figure out where in open_mfdataset/auto_concat the attributes are lost.

[pwolfram]

@shoyer, the issue is at merge of xarray/core/merge.py. Essentially the lines

variables, coord_names, dims = merge_core(dict_like_objects, compat, join)
merged = Dataset._construct_direct(variables, coord_names, dims)

drop the attributes. The solution is to do something like a merge of the attrs, e.g., a merge of the OrderedDicts. I've started a solution at #1038 if you want to take a preliminary look.

[stale]

In order to maintain a list of currently relevant issues, we mark issues as stale after a period of inactivity
If this issue remains relevant, please comment here; otherwise it will be marked as closed automatically

[jhamman]

This should have been closed by #1038.