NotImplementedError: unable to open file: libtensorflow_io.so

[dgoldenberg-audiomack]

Running on AWS in EMR, bootstrapping as follows:

pip3 install --user tensorflow==2.4.0
pip3 install --user tensorflow_recommenders==v0.4.0
# Nightly fixes https://github.com/tensorflow/io/issues/1254 (has support for strings in input parquet).
pip3 install --user tensorflow-io-nightly

Seeing a library load error as below. The code is basically like this:

    def load_dataset(ds_name, files, columns):
        dataset = tfio.IODataset.from_parquet(files[0], columns=columns)
        for file_name in files[1:]:
            ds = tfio.IODataset.from_parquet(file_name, columns=columns)
            dataset = dataset.concatenate(ds)
        return dataset

I'm loading a parquet file like this: Loading s3://my-bucket/dir1/dir2/part-00000-6f1f5a9d-95ac-462c-b148-fcb9404d6972-c000.snappy.parquet

Error:

Traceback (most recent call last):
...
(my code) -- line 236, in load_dataset
  File "/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/io_dataset.py", line 275, in from_parquet
    filename, columns=columns, internal=True
  File "/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/parquet_dataset_ops.py", line 30, in __init__
    components, shapes, dtypes = core_ops.io_parquet_readable_info(
  File "/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/__init__.py", line 88, in __getattr__
    return getattr(self._load(), attrb)
  File "/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/__init__.py", line 84, in _load
    self._mod = _load_library(self._library)
  File "/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/__init__.py", line 71, in _load_library
    + "{}, from paths: {}\ncaused by: {}".format(filename, filenames, errs)
NotImplementedError: unable to open file: libtensorflow_io.so, from paths: ['/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/libtensorflow_io.so']
caused by: ['/home/hadoop/.local/lib/python3.7/site-packages/tensorflow_io/core/python/ops/libtensorflow_io.so: undefined symbol: _ZNK10tensorflow10FileSystem8BasenameEN4absl14lts_2020_02_2511string_viewE']

[yongtang]

@dgoldenberg-audiomack With PR #1309, the tensorflow-io-nightly has been switched to depending on tf-nightly. Since there are some changes in tensorflow core repo, tf-nightly's API might be slightly different from tensorflow 2.4.0.

For your specific issues, I think if you use a specific tensorflow-io-nightly version earlier than 2021/02/10:

pip3 install --user tensorflow-io-nightly=<earlier than 2021/02/10>

The issue will be resolved for your case for now.

[dgoldenberg-audiomack]

@yongtang Thanks for your fast response on this one. Oddly, I pushed my configuration back to tensorflow_io==0.17.0 from nightly. And with that, my parquet reads appear to be working. They looked broken per #1254. Somehow, I'm not seeing the #1254 error now. I'm thinking to run on 0.17.0 for now and if something breaks with parquet reading, then use your recommendation and switch to tensorflow-io-nightly=<earlier than 2021/02/10>.

[dgoldenberg-audiomack]

@yongtang Hi, I'm looking to run some code with Tensorflow v.2.3.1.

What version of tensorflow io would you recommend that would be compatible with 2.3.1?

Also, any way to get a hold of a compatible version which would have this fix for parquet processing?

When installing tensorflow-io with pip, any way to cause it not to install tensorflow? I want to make sure I'm using the pre-installed TF 2.3.1 and not change anything there.

Thanks.

[yongtang]

@dgoldenberg-audiomack You can find version compatibility on:
https://github.com/tensorflow/io#tensorflow-version-compatibility

We normally don't do patch for past releases due to the resource limitation. We can consider it if there is a great need (or for security/vulnerability reasons).

[dgoldenberg-audiomack]

Thank you for that pointer, @yongtang.

Specifically for TF 2.3.1, it was released on Sep 24, 2020. Looking at that table, I see

0.16.0	2.3.x	Oct 23, 2020
0.15.0	2.3.x	Aug 03, 2020

I assume I'll want to run TF IO 0.16.0.

So without a patch to 0.16.0, there is no support for strings in Parquet, which I would classify as a great need. Could we please have this patch generated? Thank you.

[yongtang]

there is no support for strings in Parquet

@dgoldenberg-audiomack Do you know the specific commit that fixed this issue? To release a patch we will need the following:

1. Create a R0.16 branch
2. Cherry-pick the specific commit to the branch
3. Update the version to 0.16.1
4. Build on GitHub CI and, if everything passes we can push to pypi.org

[dgoldenberg-audiomack]

@yongtang Hi, I don't know, I was just going by your earlier comment

@dgoldenberg-audiomack With PR #1309, the tensorflow-io-nightly has been switched to depending on tf-nightly. Since there are some changes in tensorflow core repo, tf-nightly's API might be slightly different from tensorflow 2.4.0.

For your specific issues, I think if you use a specific tensorflow-io-nightly version earlier than 2021/02/10:

pip3 install --user tensorflow-io-nightly=<earlier than 2021/02/10
The issue will be resolved for your case for now.

[yongtang]

@dgoldenberg-audiomack The issue fixed by #1309 is the API compatibility issue which is different from the comment of there is no support for strings in Parquet,. Can you explain a little more about the "strings support" you are referring to in the comment?

[dgoldenberg-audiomack]

@yongtang My goal is to be able to read parquet into tf datasets without any issues.

We had started this discussion here: #1254.

You had added PR #1262 for the fix to that issue.

If I'm using this:

pip3 install --user tensorflow==2.4.0
pip3 install --user tensorflow_recommenders==v0.4.0
# Nightly fixes https://github.com/tensorflow/io/issues/1254 (has support for strings in input parquet).
pip3 install --user tensorflow-io-nightly

then the question is, will I be able to? TF 2.4.0 or 2.4.1, TFRS v0.4.0. What version of TF IO should I use? so that I have the fix for #1254 but also don't get the error described in this ticket:

libtensorflow_io.so: undefined symbol: _ZNK10tensorflow10FileSystem8BasenameEN4absl14lts_2020_02_2511string_viewE'

I'm just looking for a clean deployment. Whether any more fixes or patches are necessary, I cannot tell you.

It seems like if I do

pip3 install --user tensorflow==2.4.1
pip3 install --user tensorflow_recommenders==v0.4.0
pip3 install --user tensorflow-io
pip3 install --user boto3

then parquet is loaded with no issues. Maybe your fix for #1254 is already in TF IO latest? With this, I don't seem to get the undefined symbol, either...

[yongtang]

To release a version that includes #1262 and works with tensorflow 2.4.x, I think the easiest paths could be:

1. Release 0.17.1 with cherry-pick of Fix incomplete row reading issue in parquet files #1262 and pin to tensorflow 2.4.x, or
2. Release 0.18.0 with current master and pin to tensorflow 2.4.x, then move to tensorflow 2.5.0 for 0.19.0 release.

cc @kvignesh1420 @terrytangyuan any insight on next release?

[kvignesh1420]

@yongtang I think it's better to go with the first option: "Release 0.17.1 with cherry-pick of #1262 and pin to tensorflow 2.4.x" due to the following reasons:

• The change is minor and will not affect other API's.
• If we release 0.18.0 (current master) and pin to 2.4.x then the file-system plugins will fail to load. This is a major concern which we don't want. (for ex: the API compatibility checks fail when the current tfio master is used with tf 2.4.x. Reference)

[yongtang]

V0.17.1 has been released:
https://github.com/tensorflow/io/releases/tag/v0.17.1
https://pypi.org/project/tensorflow-io/0.17.1/

binaries:

$ sha256sum *.whl
1cbd071850901d3adc4aecbb5030e54e9c772ce71494ac8458fa67bf2ce9521a  tensorflow_io-0.17.1-cp36-cp36m-macosx_10_13_x86_64.whl
b9cf74e838ea7feab56ece6ccb58fc488bb685cc71d0b23bc12eccf68a584744  tensorflow_io-0.17.1-cp36-cp36m-manylinux2010_x86_64.whl
ae101de08bfa8b640af42ca24843c64fe53371299e1c6c56d873fe4a6bcd697c  tensorflow_io-0.17.1-cp36-cp36m-win_amd64.whl
b04ed67434d2de57451fcf7b9a1897e7d9cdb8d308ef16608f8d51167b950aad  tensorflow_io-0.17.1-cp37-cp37m-macosx_10_13_x86_64.whl
00f8a8a1d7561013be4ec0f8eb5fc1ad23ac8864e025aa5164c33c7f9a4c9d3d  tensorflow_io-0.17.1-cp37-cp37m-manylinux2010_x86_64.whl
39f46862eda5b46b98b59da86ca0a5fdc1c2296f58ecd9be57e770ac722cb164  tensorflow_io-0.17.1-cp37-cp37m-win_amd64.whl
fafca8ae03d52f28d2ca153079e730cf4e26a0f237c8e89578400a2033fe8700  tensorflow_io-0.17.1-cp38-cp38-macosx_10_13_x86_64.whl
8bcc45bb7040037161db30f917935f5ea36d10dd21551c7b50030bd0a395ef2d  tensorflow_io-0.17.1-cp38-cp38-manylinux2010_x86_64.whl
8343d604d14257806059fa620294565a168f495cc43b17c9ca779403714457bc  tensorflow_io-0.17.1-cp38-cp38-win_amd64.whl

[yongtang]

@dgoldenberg-audiomack With the release of 0.17.1 the issue should have been fixed. I will close this issue for now, but please feel free to re-open if the issue persists.