Serialization of pre-processing pipeline for CI/CD

[jhagege]

Hi, thanks for the great library.
I noticed in your examples you serialize the preprocessing pipeline.
Does it assume that the pip dependencies of the preprocessing classes need to be the exact same version ?
I'm trying to think how to package the inference workflow inside a single Dockerfile as part of a CI/CD pipeline.
How can I guarantee that I have a self-contained Docker image with the correct exact dependencies and the serialized pre-processing pipeline.

Thanks for any insights.

[ukclivecox]

Hi @jhagege
Can you give some more details of the examples you are referring to and what you mean by "pre-processing pipeline"?

[jhagege]

@cliveseldon , thanks for your quick answer.

I was referring to the following:
from the outlier_combiner example.

I find the pattern elegant and I'm wondering how to take it one step further.
Each model created is defined by three "artifacts":

• Source code (commit id)
• Preprocessor pipeline (transforms)
• Environment (python packages for example in a requirements.txt with their dependencies version).

I'd like to configure a CI pipeline to package all of those into some kind of an "uber-artifact", per model that is trained, so that it can provide an integrated environment for inference.

Thanks for any insights.

[RafalSkolasinski]

We are not really concentrating on training.

It seems that the best approach is to have a solid reproducible preparation of artefacts / trained models (kubeflow, dvc, pachyderm, ...) and then package these into a Docker image that you can deploy with Seldon.

Check our latest addition of model metadata: https://docs.seldon.io/projects/seldon-core/en/latest/reference/apis/metadata.html that allows one to make connection to the training source of the model.

[jhagege]

Thanks much, I'll review.