Record an experiment that another team can inspect
An experiment compares runs with different data, code, model settings, or environments. Record the dataset snapshot, split definition, code revision, container or environment version, parameters, metrics, and output artifacts. Compare runs on a consistent evaluation cohort and record the meaning and direction of each metric. Agent Platform Experiments and ML Metadata help organize runs and trace their inputs and outputs, but useful lineage still depends on recording the right information.
Two runs report F1 scores of 0.84 and 0.87, but one used a different test population. List the missing evidence needed before concluding that the second model is better.
Distinguish model versions from aliases
A registry organizes models and versions for evaluation and deployment. A version identifier identifies a specific registered version; a version alias is a mutable name that can move between versions. A default alias is used when a version is not specified for supported operations. Record the resolved version ID in a release record rather than relying only on an alias such as stable, because the alias can later point elsewhere. An alias is a reference, not evidence that a model passed your release gates.
A deployment record says only “stable.” A week later the alias moves. Explain why the record is ambiguous and list the version and evaluation evidence you would retain for the next release.
Trace the deployed system end to end
Connect the registered model version to its training run, dataset and feature definitions, evaluation results, serving container, preprocessing, endpoint configuration, and deployment time. Also record the rollback target and any downstream batch outputs. The model artifact alone may be insufficient to reproduce behavior if its feature transformation or dependencies changed. Control who can register, approve, deploy, and move aliases; log the release decision and relevant access.
A prediction changes after a deployment. List three possible changes outside the model weights and the records you would inspect to distinguish them.
Practice rollback and retention
A rollback plan should identify a known-good model version and compatible preprocessing, features, runtime, and traffic configuration. Test the process before an incident and verify the restored service with representative requests. Retain the records and artifacts needed for investigation according to the organization’s data and retention policies. A lineage graph is useful for tracing upstream causes and downstream impact, but it does not guarantee that deleted data or unavailable external dependencies can be recreated.
The new model’s error rate rises immediately after rollout. Describe how to identify the previous deployed version, restore compatible dependencies, confirm service health, and preserve the failed release’s evidence.
Why should a release record contain the resolved model version ID as well as an alias?
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.