Kubeflow Visualizations Have all Become Broken as of 0.29
See original GitHub issueWith recent release of TFX, artifacts paths have been changing and no longer work in Kubeflow. Shouldn’t the changes be communicated with the Kubeflow maintainers? They are still hard-coded to the old paths:
https://github.com/kubeflow/pipelines/blob/master/frontend/src/lib/OutputArtifactLoader.ts#L273-L274
Ideally both old and new paths should be supported.
I’m currently on Kubeflow 1.4.0 with TFX 0.30.0 and all visualizations (including evaluator given there is a baseline (though patched in master)) are broken.
As I upgraded from 0.27, and the tensorboard visualization doesn’t seem to work either as it seems to not be included in the ml-pipeline-ui-metadata.tgz
anymore for some reason.
As I just mentioned ml-pipeline-ui-metadata
: Can you also please consider moving /ml-pipeline-ui-metadata.json
out of the root folder to say /output/ml-pieline-ui-metadata.json
?
Writing to /ml-pipeline-ui-metadata.json
makes it impossible to use non docker
(which as you all know is deprecated now) kubeflow deployments:
Issue Analytics
- State:
- Created 2 years ago
- Reactions:4
- Comments:5 (1 by maintainers)
Top GitHub Comments
https://github.com/kubeflow/pipelines/pull/5999 has merged to fix this issue, expect to be included in the next KFP release. After upgrade, KFP requires users to upgrade TFX to be >0.30.0 for adopting this change.
It would be awesome to get this fixed. On another thread, it was reported that Kubeflow Pipelines 1.6.0 will fix the missing visualisations.