One limitation I'm concerned about with new Arcanist Hardpoint infrastructure is that it makes parallelizing service calls difficult. I'm generally satisfied with it from a power/extensibility standpoint and I think the complexity is manageable, but I'm uncomfortable about performance.
A 2013 series of changes starting with D5104 built FutureGraph, which may be a reasonable approach here. That stuff never moved forward because the motivation was parallelizing queries (which we generally have less to gain from) and we didn't get anywhere I was satisfied with on the call chaining API. However, the performance motivation is now more material and call chaining is less of a concern with the simpler Conduit futures.