Add model deployment to github workflow

[oindrillac]

Add model deployment to github workflow such that the user can interact with the trained model at a certain url

[oindrillac]

There are a few ways we can add a model deployment step to the workflow.
A relatively low external dependency approach that we can follow would be to create a separate Github workflow :

• Once a user trains a new model on a given repository (group of repos or org in the future), the model is saved on a user provided S3 bucket

• Create a new workflow that is triggered when a user opens a new PR.

  • How do we "ship" this new workflow? Will this be a .github/workflows/file that they copy on to their repo? Is this a bot that they configure. We can try to reuse the Thoth TTM bot process.

  • This Github Action workflow is triggered by opening of a new PR.

  • The output of this workflow is a time estimate for merging the PR, this workflow comments on the PR this is triggered from.

  • The workflow runs a notebook/script in the background which runs very similar to https://github.com/oindrillac/ttmtool/blob/main/04_model_inference.ipynb

    • It reads the trained model saved on S3
    • It reads the data from the PR in question using srcopsmetrics (so that its in the same format as required by the model)
    • The model pipeline has 2 steps : feature engineering and model inference. The model pipeline converts the data to the features and gives out an inference
    • The output of the model is commented on the open PR by the workflow

[oindrillac]

How do we "ship" this new workflow?

The end goal should be to make both workflows available by publishing them somewhere

• 1st action (model training) is an on demand action
• 2nd action (model inference) will preferably be an organization level workflow. Maybe an org level bot so that user doesnt need to run this on every repo individually?