Split data the way the world changes
Random splitting is not always representative. When predicting future events, a time-based split can reveal performance degradation that random sampling hides. Keep related examples together when multiple records belong to one person or device. Fit transformations on the training split, then apply those transformations to validation and test data.
Design a split for monthly customer transactions. Explain how placing one customer in both training and test sets could inflate the score.
Choose processing tools by workload
Use SQL when the transformations fit naturally into warehouse queries. For distributed or streaming processing, evaluate a pipeline framework suited to the volume and arrival pattern. In-memory Python is useful for smaller experiments when the dataset fits available memory. Track feature definitions, ownership, access, and freshness so another team can reproduce your work.
Specify a freshness requirement for a fraud feature and a churn feature. Explain why their serving requirements differ.
Record the whole experiment
A metric without context is hard to reproduce. Record the data snapshot, code revision, parameters, environment, artifacts, and evaluation procedure. Compare experiments using the same evaluation set and metric definitions. Apply least privilege to notebooks and storage; use managed credentials rather than embedding keys in a notebook.
Create an experiment log with a baseline and two changes. Change one meaningful variable at a time and explain each result.
A fraud feature uses chargeback information recorded a week after the transaction. What is the main issue?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.