It looks like with later versions they switched to kind of a hybrid approach (part-remote, part-local) with Hyper to reduce latency for interactivity.
> there is no sharing of the underlying data set for multiple projections across the same large data set
But that would require some kind of open standard for portability, no?
I like the approach AGGrid uses - they provide a viewport based interface that the grid uses to display data, and you can implement that interface on top of your data model - https://www.ag-grid.com/javascript-data-grid/viewport/. Unfortunately it's only available in their enterprise version, but this approach scales to both grid and chart based UIs. D3 has a bit of that flavor as well, since you can map visual attributes into your underlying data any way you'd like.
This applies to any charting library that forces you to provide both spec and unaggregated data to memory/cpu constrained clients (e.g. Javascript in the browser). This is done for implementation-simplicity (Vega, for example), but obviously doesn't scale to larger datasets.
I've implemented a system where the data part of the spec is munged in-database, and aggregated data is provided to the browser, along with hints for axes, scales, legends, etc. It requires a part of the GoG interpreter to be resident on the server-side.