Draw the end-to-end dependency graph
List the events and artifacts from source arrival through data validation, training, evaluation, deployment, and downstream consumption. Identify schedules, external systems, owners, and failure boundaries. A workflow scheduler coordinates dependencies; a data-processing engine transforms data; a training service runs training. These roles can coexist. Start with the existing architecture and operational requirements rather than assuming a single tool must perform every step.
A nightly workflow waits for a partner file, runs a warehouse transformation, retrains a model, and publishes a report. Draw the dependencies and mark which outputs must be validated before downstream tasks run.
Compare the orchestration options
Agent Platform Pipelines is designed to orchestrate ML workflows using supported pipeline definitions, components, artifacts, and metadata. Managed Service for Apache Airflow can coordinate broader heterogeneous workflows using Airflow DAGs and integrations. An organization may use one to trigger or coordinate work in the other; that introduces another boundary to monitor and secure. Compare existing expertise, artifact lineage, scheduling, operational ownership, integration needs, and cost.
An established Airflow workflow already coordinates warehouse and external-system tasks, while an ML team owns a reusable training pipeline. Describe a design that preserves ownership boundaries and how the outer workflow learns whether the inner run succeeded.
Make retries and release gates explicit
Retries are safe only when a task’s effects are understood. Use run-specific outputs, idempotent writes, or deduplication where needed; do not assume a retried notification or deployment is harmless. Validate source freshness and schema, evaluate a candidate on an appropriate cohort, and stop automatic promotion when required gates fail or results are missing. Record run identifiers and artifacts so operators can determine what completed and what must be retried.
A deployment task times out after sending a request, but the endpoint may already have changed. Explain why blindly retrying can be risky and what state the workflow should check first.
Operate the workflow, not just the happy path
Monitor task failures, retries, duration, data freshness, costs, and whether expected runs occurred. Record component versions, parameters, input snapshots, evaluation results, and release decisions. Separate identities for data access, training, and deployment where appropriate. Test missed schedules, late files, partial outputs, and rollback. Review the current product documentation for supported components, permissions, and service limits.
A source file arrives late and the previous day’s file is still present. Define a freshness gate, an alert, and a policy for preventing a stale dataset from being silently reused.
An ML training task succeeds, but a required evaluation artifact is missing. What should the orchestration do?
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.