Throughput is lower than TFRecords when there are many strings in Parquets file

[deepllz]

System information

• OS Platform: Ubuntu 18.04.5 LTS
• Docker version: 18.09.5
• GCC version: 7.5.0
• Python version: 3.6.9
• TensorFlow/PyTorch version: tf1.15.5

Willing to contribute

Yes