There is a quiescent overhead with all the objects,
For NRT objects it got improved somewhat by me a couple of days ago because I noticed we were updating and constraining the parameters every vector rather than just when a trigger came in.
For RT objects (which is now most of the things that are causing you woe) here and in FluidMLPRegressor killing sound on a server) it’s less straightforward. I’ll have a look to see if
- For these classes, the parameter updates can also be made conditional (not so sure)
- Whether the parameter updating code itself can be made less greedy. The problem is parsing strings and arrays.