Pipelines, Orchestration and CI/CT — GCP Professional ML Engineer practice questions

A KFP component fails intermittently due to GCS rate limits. Best fix?

1. Add exponential backoff retries plus a pipeline retry policy; partition writes
2. Ignore the intermittent failures, since a later scheduled run usually succeeds
3. Request a quota increase and leave the component's write pattern unchanged
4. Stop using Cloud Storage and write the component's outputs to the local disk

Answer: A — Add exponential backoff retries plus a pipeline retry policy; partition writes

A) The controlling idea is that rate limiting is a transient, retryable condition, so backoff plus spreading writes across key prefixes addresses both the symptom and the hot-spotting cause. C tempts because more headroom sounds like a fix, but per-prefix request rates still throttle a bursty writer. B tempts because runs do sometimes pass, yet flakiness silently corrupts schedules and trust. D tempts by removing the dependency, but local disk is not durable or shared.

A pipeline component writes its output to a hard-coded bucket path that is overwritten each run. What is the MOST significant consequence?

1. The component cannot be shared between two different projects
2. The pipeline consumes more storage than necessary over time
3. Step caching and lineage break because outputs are not run-scoped
4. Concurrent runs will fail with permission errors on the bucket

Answer: C — Step caching and lineage break because outputs are not run-scoped

C) Correct — orchestrators identify artifacts by run-scoped output paths, so overwriting a fixed location destroys both cache correctness and the lineage record. B) Overwriting actually uses less storage, not more; that is not the harm. A) Cross-project sharing is a configuration matter and would apply to parameterized paths too. D) Concurrent writes collide and corrupt results rather than raising permission errors.

Half-right MLOps Level 1: team has automated CT pipeline but never validates schema. Risk?

1. Silent feature shifts can poison training; add validation and gate on anomalies
2. No real risk, because the automated retraining pipeline reruns and self-corrects
3. Training completes faster, since skipping schema checks removes work per run
4. Costs fall, because validation is the most expensive part of a training pipeline

Answer: A — Silent feature shifts can poison training; add validation and gate on anomalies

A) The controlling idea is that continuous training amplifies upstream data errors: without schema and distribution validation, a malformed feed is learned rather than caught. B tempts because frequent retraining sounds self-healing, but it retrains on the bad data faster. C tempts because skipping a step is literally faster, yet that speed is what allows the corruption through. D tempts by framing validation as costly, but it is cheap relative to a bad model in production.

MLOps Level 1 adds:

1. Manual scripts handed between the data scientists and the operations team
2. No automated testing of the components or of the training code itself
3. Manual deployment of the trained model through the console by an operator
4. An automated, parameterized pipeline retraining continuously on new data

Answer: D — An automated, parameterized pipeline retraining continuously on new data

D) The controlling idea is that Level 1 introduces continuous training: the training workflow becomes an automated, parameterized pipeline that reruns on fresh data. A tempts because scripts still exist, but manual handoff is the defining trait of Level 0. C tempts because deployment may indeed still be manual here, yet that is what carries over, not what is added. B tempts because testing gaps persist, but absence of tests describes the prior level rather than the new capability.

A team wants pipeline artifacts versioned with code. Best practice?

1. Pin component image digests and pipeline code in Git; version the artifacts
2. Keep the artifacts on local disk on the workstation that submitted the run
3. Email the trained artifacts to the team so each member keeps their own copy
4. Track artifact names and storage locations by hand in a shared spreadsheet

Answer: A — Pin component image digests and pipeline code in Git; version the artifacts

A) The controlling idea is that reproducibility requires the code, the exact component images, and the artifacts to be versioned together and linked by recorded lineage. B tempts because local files are immediately available, but a workstation disk is volatile and invisible to the team. D tempts because a spreadsheet is easy to start, yet it drifts from reality and records nothing automatically. C tempts as a sharing mechanism, but email is not a versioned store of any kind.

Right feature wrong stage: Cloud Composer is best when:

1. You need cross-system orchestration over ETL, ML, and BI with Airflow operators
2. You need ML artifact lineage recorded automatically across every training step
3. You only need one simple recurring cron trigger for a single nightly job
4. You run distributed model training and want the scheduler to execute it

Answer: A — You need cross-system orchestration over ETL, ML, and BI with Airflow operators

A) The controlling idea is that Composer's strength is heterogeneous orchestration: Airflow operators reach many systems, so it fits workflows spanning data, ML, and reporting. B tempts because Composer does track task state, but ML-native lineage belongs to the managed pipeline service. C tempts because Composer can run on a schedule, yet standing up an Airflow environment for one cron job is heavy. D tempts because training can be triggered from it, but Composer orchestrates rather than executes training.

A pipeline should register a newly trained model only when it beats the incumbent on the holdout set. The BEST implementation is:

1. Always register the model and let a reviewer delete bad versions later
2. Register only models whose training loss fell below a fixed constant
3. Register the model and set its traffic weight to zero on the endpoint
4. A conditional gate comparing the candidate's metric to the incumbent

Answer: D — A conditional gate comparing the candidate's metric to the incumbent

D) Correct — an automated comparison against the incumbent is the standard model-validation gate and keeps the registry meaningful without human latency. A) Registering everything makes the registry noisy and relies on humans to catch regressions. C) Deploying at zero weight still promotes an unvalidated artifact and confuses the deployment story. B) Training loss is not comparable across data versions and says nothing about generalization relative to the incumbent.

A pipeline must retrain only when data drift is detected. Best trigger?

1. A fixed daily schedule that retrains whether or not the data has shifted
2. No trigger at all, relying on the endpoint to adapt as traffic changes
3. A manual retraining request filed after an analyst reviews weekly reports
4. A monitoring alert to Pub/Sub, then Cloud Run, which submits a pipeline

Answer: D — A monitoring alert to Pub/Sub, then Cloud Run, which submits a pipeline

D) The controlling idea is event-driven retraining: the monitoring job detects drift, publishes an alert, and a lightweight service submits the training pipeline run. A tempts because schedules are automation, but a fixed cadence both misses drift between runs and wastes compute when nothing changed. C tempts because human review adds judgment, yet it is not a trigger and does not meet the requirement. B tempts only if one assumes models self-correct, which they do not.

A pipeline re-runs after only its evaluation step changed. Training takes six hours and its inputs and code are identical. Which pipeline feature saves the most time?

1. Step-level caching that reuses the previous training output
2. Distributed training to shorten the six-hour training step
3. Preemptible workers to lower the training step's compute cost
4. A retry policy with backoff on the training step

Answer: A — Step-level caching that reuses the previous training output

A) Correct — when a step's inputs and definition are unchanged, caching reuses the prior artifact and the six hours are skipped entirely. B) Distributed training reduces runtime but still recomputes work whose result already exists. C) Preemptible workers address cost rather than elapsed time, and they can lengthen runs on preemption. D) Retries help with flaky failures; nothing here failed.

Continuous Training (CT) vs Continuous Integration (CI) vs Continuous Delivery (CD) — which belongs to MLOps Level 2?

1. All three: CI builds and tests, CD deploys pipelines and models, CT retrains
2. Only CT, because Level 2 is defined solely by automated continuous retraining
3. Only CD, because Level 2 is defined by automated delivery of models to serving
4. None of them, since Level 2 describes manual experimentation with scripts

Answer: A — All three: CI builds and tests, CD deploys pipelines and models, CT retrains

A) The controlling idea is that Level 2 is full automation: the pipeline itself is built and tested by CI, delivered by CD, and executes continuous training on new data. B tempts because CT is the headline capability, but CT alone already arrives at Level 1. C tempts because delivery automation is visible, yet it omits both the build and the retraining halves. D tempts by describing Level 0 and contradicts the premise of the maturity ladder.

What role does Cloud Build play in an ML CI/CD setup?

It runs the build pipeline on source commits: executing unit and component tests, building and pushing training and serving container images to Artifact Registry, compiling the pipeline definition, and submitting the pipeline run or triggering deployment. It is the automation surface that turns a commit into a reproducible pipeline execution.

Describe the three levels of MLOps maturity.

Level 0 — manual, script-driven training and handoff to deployment, with infrequent releases • Level 1 — an automated training pipeline enabling continuous training on new data, with a feature store and metadata tracking • Level 2 — full CI/CD for the pipeline itself, so pipeline code changes are tested, built and deployed automatically and multiple models are managed at scale.

Vertex AI Metadata?

ML metadata store — captures runs, parameters, metrics, artifacts. Foundation for lineage.

What does ML Metadata record for a pipeline run, and what is it used for?

It records the artifacts produced and consumed, the executions that produced them, and the contexts grouping them into runs. It is used to trace which dataset and code produced a deployed model, compare runs, and satisfy audit and reproducibility requirements without inspecting logs.