Feature Request: Support reading large batch of tabular data from Parquet files efficiently.

[2sin18]

User Story

As a recommender system engineer, I want to read large batch of tabular data from Parquet files efficiently, so that training performance of large deep recommenders can be improved.

Detailed requirements

• It should be easy to work with existing Dataset based data pipeline.
• It should be optimized for extra large batch size, and utilize features of Parquet format, e.g. column selection, batch reading , and row group filtering.
• It should be compatible with vanilla TensorFlow >= 1.14 < 2.0 .

API Compatibility

• Only new APIs should be introduced.

Willing to contribute

Yes