Default behaviour when no data is provided for a variable

[lbesnard]

It looks like it is not possible to write a netcdf file without any values attached to a variable.
see
https://github.com/aodn/aodn-netcdf-tools/blob/master/ncwriter/template.py#L410

Also I don't think the keyword "value" should be used, as it is used by default by python dictionary object. Maybe "var_values" could be used or something else.
The issue with using "value" is that
asattr(template.variables["VAR_TO_TEST"]['attributes'], 'values') always returns true even if there aren't any values attached

[mhidas]

It looks like it is not possible to write a netcdf file without any values attached to a variable.

Is that something you would want to do?
We could write a variable filled with fill values by default, I guess. However unless there's a good use case requiring that, I think this would indicate the values were not set by mistake, so the code should raise an exception.

Also I don't think the keyword "value" should be used, as it is used by default by python dictionary object.

There is a distinction between the values attribute of the dictionary object (which is a method), e,g,

In[7]: template.variables['TEMP'].values
Out[7]: <bound method OrderedDict.values of OrderedDict([(u'dimensions', [u'TIME', u'DEPTH']), (u'type', u'float32'), (u'attributes', OrderedDict([(u'standard_name', u'sea_water_temperature'), (u'units', u'degC'), (u'valid_min', 0.0), (u'valid_max', 42.0)]))])>

... and the item in the dictionary corresponding to the "values" key, e.g.

In[17]: template.variables['TEMP']['values']
Out[17]: array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

But you're right that it could be a bit confusing. My preferred alternative names are "array" or "data".

[ocehugo]

It looks like it is not possible to write a NetCDF file without any values attached to a variable. see https://github.com/aodn/aodn-netcdf-tools/blob/master/ncwriter/template.py#L410

It's completely fine, because if no values are found the operation is :

ncobj['var'][:] = None

this will fill the array with NaNs. Maybe this is not the "best" though and it's best to leave the values as empty (save space).

Also I don't think the keyword "value" should be used, as it is used by default by python dictionary object. Maybe "var_values" could be used or something else.
The issue with using "value" is that
asattr(template.variables["VAR_TO_TEST"]['attributes'], 'values') always returns true even if there aren't any values attached

Yeap, disambiguation is required because of dicts. I vote for data.

[ggalibert]

The use case of a variable with no value is mostly to group some metadata together when that is relevant under a common container variable. See for example https://www.nodc.noaa.gov/data/formats/netcdf/v2.0/timeSeriesOrthogonal.cdl where:

        int platform_variable; //............................................ RECOMMENDED - a container variable storing information about the platform. If more than one, can expand each attribute into a variable. For example, platform_call_sign and platform_ncei_code. See instrument_parameter_variable for an example.
                platform_variable:long_name = "" ; //........................ RECOMMENDED - Provide a descriptive, long name for this variable. 
                platform_variable:comment = "" ; //.......................... RECOMMENDED - Add useful, additional information here.
                platform_variable:call_sign = "" ; //........................ RECOMMENDED - This attribute identifies the call sign of the platform.          
                platform_variable:ncei_code = ""; //......................... RECOMMENDED - This attribute identifies the NCEI code of the platform. Look at http://www.nodc.noaa.gov/cgi-bin/OAS/prd/platform to find if NCEI codes are available.          
                platform_variable:wmo_code = "";//........................... RECOMMENDED - This attribute identifies the wmo code of the platform. Information on getting WMO codes is available at http://www.wmo.int/pages/prog/amp/mmop/wmo-number-rules.html          
                platform_variable:imo_code  = "";//.......................... RECOMMENDED - This attribute identifies the International Maritime Organization (IMO) number assigned by Lloyd's register.

[mhidas]

@ggalibert Ok, in that case the variable value is not important, but you could still set it to something. I think that's still a bit different from writing a file without explicitly setting the variables values to something (even if just arbitrary values, fill values or NaNs).

[mhidas]

What I'm saying is, if you just do this:

template = DatasetTemplate(dimensions={'X': 10},
                           variables={'X': {'type': 'float', 'dimensions': ['X']}}
                           )
template.to_netcdf('test2.nc')

You've probably just forgotten to set template.variables['X']['values'], so the code should raise an exception.

What does everyone think?

If you actually want it to write the variable with just fill values, all you have to do is add something like this before writing the file:

template.variables['X']['_FillValue'] = -999.
template.variables['X']['values'] = np.full(10, -999.0)

[lbesnard]

I don't see why adding this constraint to NetCDF file which is not compulsory in the NetCDF format. We have to keep in mind this tool could be used by others

[ggalibert]

You've probably just forgotten to set template.variables['X']['values'], so the code should raise an exception.

I agree, this would force everyone to produce "neat", non unexpected files.

[mhidas]

@lbesnard I don't think it's a constraint, really. It's more a safety feature. The main use case for this code is for writing actual data to a file, not fill values. If you want to write fill values, you can, but you have to do so explicitly.

We could include a bit of a shortcut, so that setting template.variables['X']['values'] = None will translate to a variable with fill values (rather than NaNs). You could even set this as a default in your template.

[mhidas]

I've added two more commits to #7 in response to the above:

• renamed 'values' to 'data';
• raise error if 'data' not provided;
• all fill values if 'data' is None

[mhidas]

Please review and merge if happy.

[lbesnard]

@lbesnard I don't think it's a constraint, really. It's more a safety feature.

Well I disagree. The case of guillaume's example is a good one.

[mhidas]

And we're still allowing for that case. In fact I've made it even easier now. If you want to create a variable with all fill values, you can even specify it in the template. E.g. in the case of Guillaume's example, all you'd need to do is add this to the variables listed in your template:

        "platform_variable": {
            "type": "int",
            "data": null
        }

Given that this is the less common use case, I think that's an acceptable amount of "overhead" required to make it work.