Rule of thumb for choosing GPActiveSetSize in bayesopt for the unknown constraint model

I am using bayesopt with a unknown constraint to optimize a function with 10 parameters. I have the following three questions:

1. When the number of feasible points the algorithm has visited is too small (that is, in the first several iterations almost all the points sampled are infeasible), the algorithm will keep on ramdomly sampling points until enough feasible points have been visited to buid a GP upon. I wonder why the algorithm does a 'random' search in this case. Shouldn't it be looking at the error GP model so that it will avoid sampling near a previously visted infeasible point?

2.In the unknown constraint case, there are two GPs: one modelling the cost and the other modelling errors. Does GPActiveSetSize refer to both or the objective-GP only? When the algorithm has visited more and more points, the time to select next point grows even faster than the time to fit the models. I am curious why: shouldn't the time to select next point be the same since it samples a fixed amount of points from the acquisition function regardless of the number of iterations?

3. Is there a rule of thumb for setting GPActiveSetSize when optimizing 10 parameters? I believe the default 300 is too small for such a high-dimensional setting, but a large GPActiveSetSize will slows down my algorithm a lot.

Thanks!