Julien Schueller

@josephmure are you done reviewing ?

yes, I expected the penalty of the two CompositeDistribution being higher maybe with more complex functions / antecedent ?

I was trying to understand the difference in kriging results that can propagate to a difference at even the first digit of kriging hyperparameters

> If this PR is supposed to remove the model based parametrization of the meta-model algorithms, then I don't understand why we would have to specify both an output sample...

> Ok but why do we need to actually store the model? The setModel() accessor could simply emit a warning, and the getModel() accessor could build a DatabaseFunction on demand,...

Indeed it fixes #1473

yes, the estimator uses an hybrid optimization/method of moments so it assumes nu>2, it works better by MLE: ``` factory = ot.MaximumLikelihoodFactory(ot.Student()) factory.setOptimizationAlgorithm(ot.Cobyla()) distribution = factory.build(sample) print(distribution.getParameter()) ``` [0.501496,0.00132358,0.996152]

Here its not clear that Cobyla is best, I dont have anything better than falling back to MLE if nu is too close to 2. ``` if (nu < 2.01)...

@WardF it is configured from https://github.com/Unidata/netcdf-c/blob/master/netCDFConfig.cmake.in

hdf5 logic is complicated https://github.com/Unidata/netcdf-c/issues/877