Question: Does h5py still find "compound" dtypes useful?

[seberg]

I am wondering if h5py is still using "compount"/"union" dtypes in h5py. That is dtypes which are a normal (non void) dtype, but additionally have fields, such as:

dt = np.dtype(("S8", np.dtype([("a", "f4"), ("b", "f4")])))

Which is a bytes dtype, but additionally supports field access via the struct dtype.

The way these are constructed in NumPy always seemed very shaky to me, and it also looks always a bit like a source of complexity that might be buggy. So I am just curious if anyone thinks these are useful enough (as opposed to asking users to use arr.view(structured_dtype))

Considering that h5py seems to at somepoint have used it, I now guess it is useful enough, but wanted to ask.

[takluyver]

I didn't know that was possible until now, and I don't know what h5py ever used it for. Regular compound dtypes are supported, but as far as I know, the data type in HDF5 can either be compound (using fields) or a string, not both at once.

I can't guarantee that it's not lurking in some odd corner of the codebase, though. And I obviously don't know if someone might have found 'one weird trick' that hinges on giving such a dtype to h5py. 🙂