You have recently trained a scikit-learn model that you plan to deploy on Vertex AI. This model will support both online and batch prediction. You need to preprocess input data for model inference…

Using a prebuilt scikit-learn prediction container typically assumes input data is already in the expected format for the model and does not easily accommodate custom preprocessing logic without further modification or a custom handler.

A Custom Prediction Routine (CPR) allows you to define a `Predictor` class with custom preprocessing logic, model inference, and post-processing. Vertex AI then uses this to build a container image, embedding your custom code and minimizing the boilerplate container setup required, which is ideal for incorporating preprocessing with minimal effort.

Creating a fully custom container requires writing a complete Dockerfile and serving script, which involves significantly more code and effort compared to leveraging a Custom Prediction Routine that abstracts much of the containerization boilerplate.

This choice is a duplicate of C and, similarly, involves more manual effort for container creation and setup than using a Custom Prediction Routine for embedding custom logic.