InferenceData.to_netcdf and Arviz.from_netcdf be able to take file objects or buffers

Most libraries that allow writing or reading files also allow reading and writing to a python file object or an IO file-like object (eg. BytesIO). Currently Arviz requires writing these files to disk which creates an awkward dance if one simply wants to upload to s3, for instance. This may not seem like a big deal, but the files can be large, you have to remember to delete them off disk after you’re done with them, etc.

Currently, to upload to S3, I have to do the following:

with tempfile.NamedTemporaryFile() as fp:
    inference_data.to_netcdf(fp.name)
    fp.seek(0)
    s3.Bucket("bucket-name").upload_fileobj(
        fp, "my-s3-file-key"
    )

Would be much easier to do this (without writing to disk)

with BytesIO() as buffer:
    inference_data.to_netcdf(buffer)
    s3.Bucket("bucket-name").upload_filobj(buffer, "my-s3-file-key")

Even better, I’d suggest using s3fs which is what Pandas does and support s3:// “protocol” so:

inference_data.to_netcdf("s3://bucket-name/my-s3-file-key")

However this last suggestion is not specifically related to the more important ability to write/read from a buffer or file object.

[OriolAbril]

I think this could be added upstream to xarray (not sure to what extent it is possible and docs are missing and to what extent it is not possible), https://github.com/pydata/xarray/issues/4122 issue seems related. After all, we are basically calling xarray.Dataseto_netcdf to do all the magic.

This other issue https://github.com/pydata/xarray/issues/2995 also looks related and it could make sense to look into zarr too.

[OriolAbril]

https://twitter.com/dopplershift/status/1286415993347047425 this could be helpful, in addition to the issues linked above