Define the pipeline as a sequence of contracts
For a weekly churn model, start with an immutable snapshot of eligible, fully labeled records. Validate the schema and data quality, prepare point-in-time features, train a candidate, evaluate it against a current baseline, and publish a decision record. Each step should declare its inputs, outputs, version, and failure behavior. A successful training task is only evidence that training ran; it says nothing about whether the candidate meets release requirements.
Draw these six steps and list one artifact produced by each. Identify where the run must stop if the label column is missing and where an existing deployed model remains the serving baseline.
Gate data and models independently
Data validation can check schema, missingness, freshness, label maturity, and unexpected values before training. Model validation should compare a candidate and baseline on the same representative evaluation cohort, then check quality, important slices, latency, and operational constraints. Define thresholds and insufficient-sample behavior before seeing results. If a required gate fails or cannot be evaluated, retain the evidence and prevent automatic promotion rather than silently treating the missing result as a pass.
The candidate improves overall recall from 0.76 to 0.80, but recall for a required segment falls from 0.70 to 0.55 and the minimum is 0.65. The overall improvement does not satisfy the segment gate; block automatic promotion and investigate, including the segment sample size.
Make runs reproducible and caching safe
Record the pipeline and component versions, immutable image digest, parameters, source revision, data snapshot identifiers, model artifact, evaluation results, and release decision. Agent Platform Pipelines records artifact and parameter metadata for lineage. Its execution cache uses the component specification and declared inputs; an external table that changes behind the same undeclared reference can make a reused result stale. Make mutable dependencies explicit through versioned snapshots or parameters, or disable caching for that step when appropriate.
A feature-building step reads a mutable table by name while its declared inputs stay unchanged. Explain why a cached result might be stale and propose one way to make the data version part of the step’s input.
Separate retraining from promotion
A schedule, newly available labels, or a monitoring alert can trigger investigation or a training run. An input-drift alert is not proof that a candidate will improve outcomes. Require matured labels, validation, and a promotion decision before changing production. Track failures, retries, permissions, and costs; make retryable steps idempotent or write to run-specific destinations. Keep the previous model and compatible dependencies available for a tested rollback.
A drift alarm fires on Monday, but the outcome labels arrive after 30 days. Describe what can be investigated immediately, what data can be used for training, and which evidence is needed before promotion.
A training run succeeds but the candidate fails a mandatory segment-quality gate. What should happen?
Keep exploring
- Agent Platform Pipelines and artifact lineage ↗
- Pipeline execution caching ↗
- MLOps continuous training guidance ↗
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.