[FEA] Add write_json method to cudf/io/json.hpp

[dagardner-nv]

Is your feature request related to a problem? Please describe.
Currently the only method for serializing a cudf dataframe to json requires acquiring the Python GIL.

cudf/io/json.hpp contains a read_json method, and the corresponding csv.hpp header contains both read_csv and write_csv methods.

Describe the solution you'd like
Add a write_json method which is callable from C++ and does not require the Python GIL.

Describe alternatives you've considered
Currently we have to use the Python API from C++:
https://github.com/nv-morpheus/Morpheus/blob/branch-22.08/morpheus/_lib/src/io/serializers.cpp#L109

Additional context
This would be a big performance win for us to not require the GIL in our serialize stage.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[GregoryKimball]

Thank you @dagardner-nv for sharing this request. Currently our main JSON focus is on reading arbitrarily-nested data types (Nested JSON reader).

Writing JSON is complex because there are several common ways to represent tabular data in JSON. Would you please review the orient argument in pandas.read_json and share which convention would match your use case?

EDIT: I see from your link that you are writing JSON Lines as records. Thank you!

[dagardner-nv]

@GregoryKimball thanks, and yes for the purposes of Morpheus most of the time by the time we get around to writing it the data would be flat.

I wonder if it could be possible to do something like:

if is_simple(df):
    use_cpp_impl(df)
else:
    # Use current impl

[GregoryKimball]

Hello @chinmaychandak, would this to_json feature be useful for your work in Kratos?

[chinmaychandak]

Thanks for pointing me to this @GregoryKimball!

Yes, to_json without acquiring the Python GIL would be super amazing for both overall performance like David mentioned as well as for streaming computations involving multiple layers of transformations and while writing serialized data (JSON Lines, mostly, in our case) back to sinks like Kafka/S3 from cuStreamz pipelines.

@jsmaupin - anything to add?

[jsmaupin]

Nothing to add from me. I agree with @chinmaychandak 's comments.