What you would like to accomplish:

Easier deployment to vertex of bigquery-ML models that use feature-store.

Currently, when training a BQML model WITHOUT using feature store, one can pass these parameters

MODEL_REGISTRY, VERTEX_AI_MODEL_ID, VERTEX_AI_MODEL_VERSION_ALIASES

and then deploy the model from the vertexAI model registry, by using just the webUI.

But once a feature store is added, one is forced to code quite a bit to add the feature store.

How this might work:

In bigquery-ML training, define a feature-view as parameter.

Then when deploying in vertexAI's webUI, use that feature store to map the inputs, then pass them to the model.

The inputs/outputs of the feature store's mapping are available if the query for training is constructed carefully

(join the BQ tables the feature-view is based on, don't re-use column names that show up in those tables).

To show how we would use it:

We run multiple models e.g. for purchase-probability-prediction and fraud detection.

We like to map e.g. the email-domain to bounce-rate, fraud-rate, frequency, etc.

These mappings are used in multiple models, and updating them weekly is of great benefit (e.g. to have data on newest OS version, latest fraud-rates, etc.).

Hence the feature store is ideal, except the deployment gets too hard.

If applicable, reasons why alternative solutions are not sufficient:

A no-code solution is very attractive to business intelligence teams, who are good at creating models, but not at writing deployable python code.

Our team is one example. And tabular models with good feature engineering (mapping by feature store, TRANSFORM feature) are our bread-and-butter!

Other information (workarounds you have tried, documentation consulted, etc):

A vertex-custom-prediction-routine is possible.

But the documentation is awful at the moment, main issues are/were:

• for the exported BQML models, the AdapterPredictor hint is not easy to find: https://cloud.google.com/vertex-ai/docs/predictions/migrate-cpr#prepare_the_source_folder_for_vertex_ai_deployment
• how to integrate TRANSFORM in particular (without feature store it is just deployed automatically)

A cloud run endpoint that adds feature-store data is possible (https://github.com/GoogleCloudPlatform/fraudfinder), but too much code as well. It does deploy the TRANSFORM it seems. But 2 endpoints (cloud run + vertex) are more costly.

Comment

If feature store was deployed automatically, we would be spending money on this already. I believe there are more teams like ours.

On the other hand, there is still a severe bug in TRANSFORM's deployment that I already reported. So we can't migrate to this yet.