What's changed: Created Professional Machine Learning Engineer Chapter 5 (Domain 5 "Pipelines": data/model validation, consistent preprocessing, MLFlow hosting, components/parameters/triggers/Cloud Build/Cloud Run, orchestration (Kubeflow Pipelines/Gemini Enterprise Agent Platform Pipelines/Cloud Composer), hybrid/multicloud, TFX/Kubeflow DSL; retraining policy, CI/CD (Cloud Build/Jenkins), Experiments on Agent Platform/Vertex ML Metadata, model/data versioning, lineage).
5.2Retraining and metadata tracking
Understand determining an appropriate retraining policy, CI/CD model deployment (Cloud Build, Jenkins), tracking/comparing model artifacts and versions (Experiments on Agent Platform, Vertex ML Metadata), hooking into model/dataset versioning, and model and data lineage.
A model is not done once built. Retrain as it degrades, deliver safely via CI/CD, and track every artifact for reproducibility and auditability.
5.2.1Retraining policy and CI/CD
Models degrade (drift) over time, so set a retraining policy: periodic (scheduled) retraining or monitoring-triggered retraining (on performance/drift), chosen by cost and degradation speed. Test, validate, and deploy retrained models via CI/CD (Cloud Build or Jenkins), automated as continuous training (CT). Switch safely with A/B/canary. Map "retrain on degradation = monitoring-triggered retraining" and "automate model test/deploy = CI/CD (Cloud Build/Jenkins)."
5.2.2Metadata, versioning, and lineage
For reproducibility and audit, track all artifacts. Compare experiment params/metrics with Experiments on Agent Platform, and record pipeline-produced artifacts/executions/lineage with Vertex ML Metadata. Hook model and dataset versioning together so you can trace lineage—"which data, which code, which parameters produced this model." This enables root-cause analysis on incidents and regulatory compliance. Map "record artifacts/executions/lineage = Vertex ML Metadata" and "trace a model's origin = lineage."
Common: requirement → means. E.g., "auto-trigger retraining on model degradation" = monitoring-triggered retraining (CT); "test and auto-deploy retrained models" = CI/CD (Cloud Build/Jenkins); "record pipeline artifacts/executions/lineage" = Vertex ML Metadata; "trace what produced this model" = lineage; "compare experiment params/metrics" = Experiments on Agent Platform.
Watch the mix-ups: (1) Experiments on Agent Platform (record/compare runs) vs Vertex ML Metadata (record pipeline artifacts/lineage) are complementary. (2) Retraining policy trades cost vs degradation speed—more frequent is not always optimal. (3) Without lineage you cannot audit or root-cause—always record metadata.
5.2.3Section summary
- Retraining = scheduled or monitoring-triggered (cost vs degradation); CI/CD for auto test/deploy
- Record artifacts/executions/lineage = Vertex ML Metadata; compare runs = Experiments on Agent Platform
- Hook model/data versioning to trace lineage (audit/root-cause)
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. To auto-trigger retraining when model performance degrades (drift), which policy is best?
Q2. To build a pipeline that auto-deploys retrained models through test/validation, which is best?
Q3. To record pipeline artifacts, executions, and lineage to trace later, which is best?
Q4. For audit, to trace "which dataset, code, and parameters produced this model," what do you need?
Q5. Which is the best way to think about designing a retraining policy?
Keep track of your progress
The full study guide is free to read. Sign up free to practice with the question bank, track what you have read, review your mistakes, and highlight passages.