We picked perspective over datashader and others for most basic charts & cross-filtering in louie.ai. It checks a LOT of boxes for intense tools. We do GPU server-accelerated things for bigger viz (graphistry), and for frontend interactive (sub 1M rows), perspective is great. Likewise, for databrushing, we also occasionally swap in plotly for perspective.

We see a lot of scenarios like enterprise and web, and so having predictable scaling is important. Perspective is a useful tool here and offloads a lot of bursty server pressure by having strong clientside perf. You don't want 1 data scientist bringing down everyone else, so doing server-accelerated compute takes a lot more thinking. It took us years to smooth out serverside shared GPU compute with Graphistry, and we're still doing a lot more.

Otoh, for both perspective and especially datashader... the aesthetic and usability are abrasive & brutal, both for static viz + interactivity UX. I don't blame the authors, esp given the project roots. At the same time, an area I hope they will be able to afford investment in and will grow the community. Finance folks are used to this because 80s/90s quants disallow change now that they run everything, but it's a problem for most other people. We couldn't use either for core Graphistry, and are slowly working around this in Perspective for Louie.AI.