Right, you can definitely use HoloViews + Datashader to display a small viewport of a much larger dataset, with only the currently visible portion passed into the browser. With the right file format and using a Dask array or dataframe with regular sampling along N, it should even be feasible to avoid ever loading into Python memory the portion of the data that you aren’t using, lazily leaving that on disk until you pan over to that range (though I haven’t actually tried that).
Then once you have displayed the values in a window, Bokeh will be working in the global N coordinate system even though only W samples are currently displayed, and so if you use the annotation and drawing tools to collect inputs from the user about how you want those values to change, it should be easy enough to take those inputs, change the underlying large data structure, and have the results update.
I think this is a case where we should create examples in HoloViews or hvPlot showing how to do this, and in particular how to enforce a maximum size on w so that no one can zoom out and thus trigger the entire set of data to display (with potentially disastrous consequences). But it’s definitely nearly there already, and at most needs some tweaking.