Installing kubeflow-training breaks import for other kubeflow packages (katib, fairing, etc.)

[cakeislife100]

Installing kubeflow-training (at the user-level) and then trying to import katib (previously installed at the system level) leads to the error:

>>> from kubeflow import katib
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: cannot import name 'katib' from 'kubeflow' (/home/jovyan/.local/lib/python3.8/site-packages/kubeflow/__init__.py)

To repro:

1. Install kubeflow-katib at the system level using pip install kubeflow-katib
2. Install kubeflow-training at the user level using pip install --user kubeflow-training
3. Try to import katib

I believe the fix should be to revert __init__.py back to what it was in this commit, unless there is a specific reason this was removed?

__path__ = __import__('pkgutil').extend_path(__path__, __name__)

[gaocegege]

cc @alembiewski

[gaocegege]

Thanks for the issue. I am not familiar with the python client, maybe @alembiewski could help us.

6523d8d#diff-b7637017217263593c31539d809c79ef7ddcc2018b8b454ba6bdbacdb6962aaaL1

[alembiewski]

Thanks for finding that, @cakeislife100! The fix you provided in the description looks good to me - it seems like this code was lost during the major SDK upgrade. Feel free to open a PR or let me know if you need any help with that.

[jaiganeshp]

Will this issue be resolved with v1.4.0 release? I still don't see the fix in master code. It is breaking kubeflow katib for us as well.

[cakeislife100]

Sorry for the delay, I completely lost track of this. @alembiewski do you mind reviewing?

[jaiganeshp]

Thank you for the reply @cakeislife100 . @alembiewski : Do you think it could be released with v1.4.0?

[alembiewski]

@jaiganeshp, it definitely makes sense to me to include it. cc @gaocegege and @terrytangyuan to confirm

[terrytangyuan]

Yeah I think we can include it for the next 1.4 RC release.

[terrytangyuan]

cc @johnugeorge

[jaiganeshp]

Thank you so much @alembiewski and @terrytangyuan . And thank you @cakeislife100 for the fix.

[terrytangyuan]

@johnugeorge Could you cherry-pick this #1531 to the 1.4 release branch?

[terrytangyuan]

Actually this doesn’t have to be part of the release of the repo. We can just release the Python SDK independently as the versioning is different. https://pypi.org/project/kubeflow-training/#history

@alembiewski Would you like to release a new version of the SDK once 1.4 is officially released?

[alembiewski]

@terrytangyuan, sure!

We can just release the Python SDK independently as the versioning is different

But the package version should be 1.4, right? To match the operator version?

[terrytangyuan]

Yes

[johnugeorge]

Sorry that I missed this. As @terrytangyuan mentioned, this can go in the separate sdk release. We can close this issue once sdk is released

[terrytangyuan]

We are releasing 1.4 now.

@alembiewski Would you be able to help us released the new SDK with the fix?

[alembiewski]

@terrytangyuan, sure! working on it

[johnugeorge]

@alembiewski Please use the release branch https://github.com/kubeflow/training-operator/tree/v1.4-branch

[terrytangyuan]

@alembiewski Thanks! I've also created an issue to track the release automation. #1540 Feel free to pick it up if that's something you are interested in working on.

[johnugeorge]

Closing this PR as #1539 and #1541 are merged