[📄 Docs] Create a `datasets` performance guide.

[dynamicwebpaige]

Brief Overview

Downloading, saving, and preprocessing large datasets from the datasets library can often result in performance bottlenecks. These performance snags can be challenging to identify and to debug, especially for users who are less experienced with building deep learning experiments.

Feature Request

Could we create a performance guide for using datasets, similar to:

• Better performance with the tf.data API
• Analyze tf.data performance with the TF Profiler

This performance guide should detail practical options for improving performance with datasets, and enumerate any common best practices. It should also show how to use tools like the PyTorch Profiler or the TF Profiler to identify any performance bottlenecks (example below).

Related Issues

• wiki_dpr pre-processing performance #1670
• Adjusting chunk size for streaming datasets #3499
• how large datasets are handled under the hood #1004
• using map on loaded Tokenizer 10x - 100x slower than default Tokenizer? #1830
• Best way to batch a large dataset? #315
• Saving processed dataset running infinitely #1911

[lhoestq]

Hi ! Yes this is definitely something we'll explore, since optimizing processing pipelines can be challenging and because performance is key here: we want anyone to be able to play with large-scale datasets more easily.

I think we'll start by documenting the performance of the dataset transforms we provide, and then we can have some tools to help debugging/optimizing them