The Research team reached out to us that the link recommendation training pipeline setup was failing and they shared the screenshots below:
We are going to investigate the cause of this and provide a solution.
Description
Description
We are going to investigate the cause of this and provide a solution.
Details
Details
| Subject | Repo | Branch | Lines +/- | |
|---|---|---|---|---|
| requirements: update packages for model training pipeline setup | research/mwaddlink | main | +7 -8 |
Related Objects
Related Objects
Event Timeline
Comment Actions
From a quick check, the link recommendation training pipeline uses conda-analytics and the last time this training pipeline was setup, conda-analytics was using Python 3.10.8 and supported: smart-open==2.2.0, wmfdata==2.0.0. At the moment conda-analytics uses Python 3.10.12 and supports smart-open==6.4.0, wmfdata==2.0.1.
Comment Actions
I have dug a little more and found that this setup is indeed failing because some package version numbers are way behind specifically these packages:
jupyter-core packaging requests attrs smart-open wmfdata
Comment Actions
Change 978663 had a related patch set uploaded (by Kevin Bazira; author: Kevin Bazira):
[research/mwaddlink@main] requirements: update packages for model training pipeline setup
Comment Actions
Change 978663 merged by jenkins-bot:
[research/mwaddlink@main] requirements: update packages for model training pipeline setup
Comment Actions
The instructions for setting up virtual environments have a caveat that encourages users to install pyicu with OS-specific packages. However, we found that when setting up virtual environments for the training pipeline, pyicu==2.9 in requirements.txt installs without any issues.
@MGerlach, and I contemplated removing these instructions since they cause confusion. Before we decided to remove them, we checked whether pyicu is used elsewhere. It turns out that it is used in a blubber.yaml file to create a docker image. Although we don't use this image for the training pipeline, we noticed that it is used to serve the Flask app that creates an API to query link-recommendation models: https://api.wikimedia.org/service/linkrecommendation/apidocs/
We have decided to leave these instructions in place, as they may help people building and using this image.
Comment Actions
To fix the link recommendation model training pipeline, I followed the steps below:
- updated the requirements that were failing
- setup both the python3.10 conda env and python3.7 env
- run the training pipeline with a small wiki xhwiki
- the pipeline run and completed successfully
- the spark jobs, wikipedia2vec, and sqlitedict run without any issues
- pushed a patch with the updated packages