SONATA Python bindings should expose the schema of the file or datatypes of attributes

[matz-e]

Currently, I have to iterate over the attribute names, extract a bogus array with an empty selection to get the schema of a SONATA file. It would be nice if there was some direct access to the column data types or the complete schema.

[mgeplf]

By schema, do you mean dtypes?

[matz-e]

Yes

[mgeplf]

Seems to make sense to me, although I like your work around. What would the C++ API look like?

[matz-e]

I think there's no one-to-one mapping. IIRC, in C++ you just request the HighFive dataset and HDF5 will convert between what you request to read, and what is on disk. Basically a lazy access to attributes with deferred reading that should let you peak at the datatype.

But one could have a C++ schema that's just unordered_map<string, HighFive::DataType> (or equivalent).

(note that this should also include data types of {source,target}_node_id)

[mgeplf]

I think there's no one-to-one mapping.

Not sure what you mean here.

But one could have a C++ schema that's just unordered_map<string, HighFive::DataType>

Is the HighFive::DataType really what you want, though? That seems to be exposing an implementation detail out of the interface, which I don't agree with. There is also a mismatch between the what the HighFive::DataType is, and what is actually returned, which an API consumer would find confusing, I think.

I guess make a PR, and we can nail the details down in that.

[matz-e]

I don't really think we need a C++ counterpart to this, going through the call chain, we convert to whatever the "customer" requests. Whereas in Python, we reflect the datatype on disk. So I would just implement a Python side of this for now.