ADD: ESGF project abstraction

[nocollier]

As part of #47 Max and I discussed no longer trying to guess the columns of the dataframe (which define the project Dataset record id) from the response itself. CMIP6 does a good job of this but the prior projects do not and it leads to a good bit of code complexity. Instead, we will just make the projects we support explicit and use an abstract base class to define how facets are used in the code. These are things like, what facet is used for the variable? It is variable in CMIP3 and 5 but variable_id in CMIP6. Technically this limits how intake-esgf can be used but not in a way that limits what any users are doing now. If a user needs another project supported, we simply add a an implementation of the project base class.

• Add project abstraction for CMIP{3|5|6} and use it to define columns in the dataframe.
• Remove the need to specify particular facets in the code, the project abstraction should serve all these needs (i.e. model groups, relaxation facets, etc)

[nocollier]

@mgrover1 See what you think about the projects.py file I added.

[mgrover1]

I like this approach @nocollier - I think docs will be critical here in terms of clarifiying the terminology and such, but the idea of using the base class, adding the not-implemented errors, and building out for CMIP3, 5, and 6 looks great.

The other approach here would be use some sort of json/yaml spec, but the use of classes seems more straightforward/error-prone.

Please let me know when you would me to give a more thorough review - but it looks like this is the approach to go with

[nocollier]

Thanks for taking an early look. Just trying not to paint myself into a corner. In reviewing the code I have realized that a lot of the functionality (automatic cell measures for example) have cmip6 specific code anyway which this patch should fix. I agree documentation about the terminology will be key, I have short descriptions started in the base class but maybe they should just be repeated for the particularizations?

[nocollier]

The more that I look into the code, the more I realize that a lot of the nice features won't work for non-CMIP6 projects. One issue is that I assumed that the control vocabulary was always present in the dataset global attributes--this isn't true. I am disabling some things for now, but will restore them in the to_dataset_dict() rework. If we can insert the facets (those that make up the master_id) into the datasets, this would fix the poor publishing.