Robert Kern

The general `arr.__reduce_ex__(pickle.HIGHEST_PROTOCOL)` is almost there if it would use `PickleBuffer` more, but in the case of `StringDType`, it's returning a list of strings for its data (which is workable,...

But for the `versioned_hdf5` case, a numpy-generic implementation may not be fit for purpose. The main thing they need to figure out is if there is an existing HDF5-serialized chunk...

Given your constraints, I think you're probably better off just testing for `StringDType` explicitly and iterating over values one by one like you do with `dtype=object`. I suspect that all...

I was thinking more of "do it twice"; i.e. do the conversion just to compute the hash then throw it away. It's quite likely faster than your current pure Python...

> This is related to something mentioned in the footnotes. I'll move that discussion to [gh-19524](https://github.com/scipy/scipy/issues/19524) and ping you. > But I suppose one solution to this problem is to...

I see `ContinuousDistribution`s as, well, distributions and not random variables, per se. Is `dist.sample(100)` sampling 100 realizations of a single RV or a realization each from 100 RVs described the...

Reassigning the `.rng` attribute creates more problems than it solves. In the middle of a function where I'm passed an opaque distribution instance, I have no idea what it has...

Where is this discussion taking place?

And where are the public meeting minutes? We don't develop scipy in private.

I would love to see links to those calls for participation. I certainly missed them, or at least missed that they entailed discussions that I wouldn't be able to opt...