Avoid error if plugins .so module is not available #1302
Conversation
This PR raises a warning instead of an error in case plugins .so module is not available, so that tensorflow-io package can be at least partially used with python-only functions. Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
| plugin_ops = _load_library("libtensorflow_io.so", "fs") | ||
| try: | ||
| core_ops = _load_library("libtensorflow_io.so") | ||
| plugin_ops = _load_library("libtensorflow_io.so", "fs") |
There was a problem hiding this comment.
should we reload plugin_ops = _load_library("libtensorflow_io.so", "fs") only?
There was a problem hiding this comment.
@burgerkingeater This part is specific to CentOS 7. In CentOS 7 we used static linking, that prompt us to combine libtensorflow_io.so and libtensorflow_io_plugins.so into one libtensorflow_io.so as static linking does not work with two .so. Notice the comment about:
Note: load libtensorflow_io.so imperatively in case of statically linking
When we have one libtensorflow_io.so file to load, due to tensorflow's limitations, it needs to be loaded with full symbols before file system symbols. So we use core_ops = _load_library("libtensorflow_io.so") before fs load.
This also means tensorflow-io on CentOS 7 can not have api backward-compatibility like other linux dist. This is a limitation that may only be resolved in CentOS 8 or higher.
This PR raises a warning instead of an error in case plugins .so module is not available, so that tensorflow-io package can be at least partially used with python-only functions. Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
This PR raises a warning instead of an error in case plugins .so module is not available, so that tensorflow-io package can be at least partially used with python-only functions. Signed-off-by: Yong Tang <yong.tang.github@outlook.com>
This PR raises a warning instead of an error in case plugins .so module is not available, so that tensorflow-io package can be at least partially used with python-only functions.
This PR is related to #1298
Signed-off-by: Yong Tang yong.tang.github@outlook.com