SIGSEGV (Address boundary error) in tf.io.gfile with TF_USE_MODULAR_FILESYSTEM=1

[lgeiger]

A SIGSEGV (Address boundary error) fault can be caused by trying to call tensorflow gfile API when a TF_USE_MODULAR_FILESYSTEM=1 is set. This can be reproduced in version 2.16.1 and 2.17.0rc0 and the latest nightly. Probably there are more versions affected, though.

The issue can be reproduced with the following one-liner:

TF_USE_MODULAR_FILESYSTEM=1 python -c 'import tensorflow as tf;tf.io.gfile.exists("gs://tfds-data/dataset_info/mnist/3.0.1/dataset_info.json")'

which crashes with

Job 1, 'TF_USE_MODULAR_FILESYSTEM=1 pyt…' terminated by signal SIGSEGV (Address boundary error)

Posting this issue here as requested in tensorflow/tensorflow#70101 (comment)