Determining the "right" instance type running Jupyter notebook in Sagemaker when reading/writing a huge parquet file?

For notebook instance it's mostly trial-and-error, at least for now. Once your model is ready to be deployed, there is the SageMaker Inference Recommender that can do automated load testing and give you recommendation on the instance size.

It's hard to give a recommendation on the notebook instance because you might test a 100MB dataset today, but choose to go with a 500GB dataset tomorrow, so the recommendations are no longer valid.

You might want to try experimenting with a smaller dataset sampled from the original big dataset, once you are confident with the model training code, use distributed training to run it on the complete big dataset.