Shared data for finite GP about stheno.jl HOT 1 CLOSED

Ah interesting use-case! Glad to hear people are trying to use it for Bayesopt.

I guess I had been implicity assuming immutability throughout Stheno -- once data is passed into a GP, I wasn't expecting it to change.

Also, I don't think the shared data has any real benefit as plotting the gp fails if you change the data and don't recompute the gp. Also, shown in the example.

It's just a performance issue I guess 🤷

I would very much prefer to keep Stheno's assumptions about immutability, so I think the best solution here is for Stheno to include some proper sequential conditioning functionality so that you don't have to think about mutating your inputs. We've already got this in AbstractGPs.jl (not tested for performance yet though) and I'm going to be moving Stheno over to use AbstractGPs so I'm hoping to get that functionality for free in Stheno (no promises on the timeline for that though).

A good solution in the mean time is, as you say, either to copy the data or to just to construct a new posterior object whenever you modify your data.

from stheno.jl.