Why are we using this kind of hacks again and again? I've mentioned in https://github.com/apache/arrow/pull/1223#issuecomment-359892433 about the downsides of importing a giant library like pytorch and tensorflow. Importing them may appear to solve some symbol collision issue, but in the long run it only makes things worse by bringing in more and more symbols from all the libraries.
I recently found that pyarrow appears to also have symbol collision issues with [horovod](https://github.com/uber/horovod/). Do you want to also add `compat.import_horovod_extension()`? I hope not. `readelf -Ws libarrow.so.10 | grep 'WEAK.*DEFAULT' | grep -v 'arrow'` is giving me many symbols that should not appear: flatbuffers, protobuf, rapidjson.... I think that's where you want to look at [ Full content available at: https://github.com/apache/arrow/pull/2329 ] This message was relayed via gitbox.apache.org for [email protected]
