Question

Your ML team has a Vertex AI Pipeline definition that trains and evaluates a fraud model. The same pipeline must run in three environments: 'dev' (small sample, cheap machine type, no deployment), 'staging' (full data, standard machine, deploy to a test endpoint), and 'prod' (full data, deploy to production endpoint after an approval step). You want to maintain a single compiled pipeline and avoid duplicating component code across environments. What is the most maintainable approach?

Accepted Answer

Define runtime parameters (e.g., data_sample_fraction, machine_type, deploy_target, require_approval) on the pipeline and pass different parameterValues in each PipelineJob submission per environment.

Answer

Compile three separate pipeline packages, one per environment, hardcoding the sample size, machine type, and deployment target into each component.

Answer

Use a single set of hardcoded values and manually edit the compiled pipeline JSON before each environment run.

Answer

Create three copies of every component with environment-specific logic embedded and select them with Python if-statements at compile time.