Thanks all.
It’s interesting to me that everyone is choosing the “data cube” model rather than the group-centric model proposed in Xarray and collections of forecasts. Indeed, that’s what I chose too. it works well for write time, but clearly has issues for some time-centric read queries.
To that end, I wrote a custom Xarray index to help with that “FMRC”-style indexing (but does not solve the performance challenges). Followup post coming soon!