Do not propagate dataset property through projection operations

[jonmmease]

This PR addresses an issue I ran into while trying to create a version of the glaciers demo using the HoloViews link_selections function.

In this demo, the kdims of a Dataset are projected using the GeoViews project_points operation. This operation returns a new Dataset with kdims that have the same names, but their values are transformed into a new coordinate system. Without this PR, the .dataset property of the resulting object will be inherited from the parent object. This breaks link_selections because you end up with fields in obj.dataset that have the same names as fields in obj, but their values are in different coordinate systems.

Along with holoviz/holoviews#4137, This PR simply disables dataset propagation through projection operations.

An alternative to disabling propagation would be to require that projections introduce new field names, but I expect that this would be a significantly breaking change

[philippjfr]

An alternative to disabling propagation would be to require that projections introduce new field names, but I expect that this would be a significantly breaking change

I think in the longer term we could make use of the Dimension.unit field to make these distinctions but for now this seems reasonable.

[philippjfr]

Hmm, can't really work out what's up with the failing tests yet.

[jonmmease]

Hmm, can't really work out what's up with the failing tests yet.

Oh, right. These tests that were updated will need holoviz/holoviews#4137 released in order to pass.

[philippjfr]

I'll try to get this merged, but do you think we could have some mechanism to register transforms which convert the coordinates back to the original coordinate system?

[jonmmease]

do you think we could have some mechanism to register transforms which convert the coordinates back to the original coordinate system?

From the perspective of link_selections, we would need a way to invert the transformation and somehow run the selection expression through that inverse operation so that the inverse-transformed expression can be meaningfully applied to the element's .dataset object.

I'm thinking more and more that we may need to make a selection a class that can produce a dim-expression on demand. That way the concept of a coordinate transform could be handled by the class, rather than trying to perform the transform on the dim expression.

[jonmmease]

I think we can close this @philippjfr, though I don't exactly remember the outcome. Did you do some work in HoloViews itself to handle projections in the pipeline?

[philippjfr]

Still need to revisit this. I'll leave it open for now.

[jonmmease]

Closing to clean up my PR history, but feel free to re-open if someone wants to pick this up again.