Task 2.4: Leveraging Gemini Enterprise Agent Platform for end-to-end ML workflows.

• AA Managed Airflow environment that runs each step

  Incorrect

  Managed Airflow is a general workflow orchestrator in an environment the team creates; it does not record ML artifacts and their lineage in ML Metadata.

• BAgent Platform Pipelines with a conditional deployment step

  Correct

  Agent Platform Pipelines runs Kubeflow or TFX pipelines serverlessly and records each run's artifacts and lineage in ML Metadata.

• CCloud Scheduler triggering a chain of Cloud Run functions

  Incorrect

  A scheduled chain of functions can run the steps but gives no pipeline definition, reproducibility or lineage.

• DA Dataflow batch pipeline that runs each step as a transform

  Incorrect

  Dataflow processes data at scale; it is not an orchestrator for training, evaluation and deployment.

Agent Platform Pipelines is the managed way to automate the ML lifecycle: each step is a component, runs are reproducible from the pipeline definition, and ML Metadata tracks lineage.

• ASpot VMs, which are discounted and need no commitment for the GPUs

  Incorrect

  Spot VMs are preemptible, which the run cannot tolerate.

• BFlex-start, which provisions the GPUs when capacity frees up for up to seven days

  Correct

  Flex-start suits short jobs of up to seven days that can wait for capacity; it is discounted and not preemptible once running.

• CA standard reservation covered by a three-year committed use discount

  Incorrect

  The discount comes from a long-term commitment, which the team wants to avoid.

• DOn-demand VMs created by hand once the training data is ready

  Incorrect

  On-demand capacity has no discount and may not be available for 64 GPUs at once.

AI Hypercomputer's consumption options trade capacity assurance, preemption and price. A short, non-preemptible job with a flexible start date is the Flex-start case.

• AAgent Platform Feature Store using the BigQuery tables as its source

  Correct

  Feature Store serves the latest feature values online at low latency directly from the BigQuery tables or views, so the same feature data feeds training and serving.

• BA query to BigQuery from the online service for every prediction

  Incorrect

  BigQuery is built for analytics, not low-latency lookups on every request.

• CA custom job that copies features into Memorystore every few minutes

  Incorrect

  This works, but it is the second pipeline and copy of the logic the team wants to avoid.

• DA Firestore collection that the training job writes features into

  Incorrect

  It adds another store and pipeline to keep consistent with the BigQuery features.

Preparing data for Agent Platform includes serving features consistently. Feature Store uses BigQuery as the source of truth and adds online serving on top of it.

• ACloud Monitoring alerts on the endpoint's CPU utilization

  Incorrect

  Infrastructure metrics show load, not changes in what the model is being asked to predict.

• BModel Monitoring with input feature drift objectives and alerts

  Correct

  Model Monitoring compares production feature distributions with a baseline and alerts when drift exceeds a threshold.

• CError Reporting on the endpoint's prediction request logs

  Incorrect

  Error Reporting groups exceptions; drifting inputs do not raise errors.

• DA weekly retraining job that replaces the deployed model every time

  Incorrect

  Blind retraining hides the signal and can make things worse; it does not detect drift.

The ML lifecycle does not end at deployment. Monitoring for skew and drift tells the team when to re-evaluate or retrain.

• AAs a batch inference job that reads from and writes to BigQuery

  Correct

  Batch inference processes a whole input source and writes results to Cloud Storage or BigQuery, with no endpoint to keep running.

• BThrough an online endpoint called once per row by a script

  Incorrect

  An online endpoint must stay deployed, and 200 million calls is the slow, costly way to score a table.

• CThrough an online endpoint with autoscaling set to a high node count

  Incorrect

  Autoscaling helps online traffic, but the endpoint still costs money between runs.

• DIn a Cloud Run function that loads the model for every row

  Incorrect

  Loading a model per row is slow and fragile at this volume.

Choosing a serving mode is part of an end-to-end ML workflow. When latency does not matter, batch inference is cheaper and simpler than online serving.

• AAutoML training on the same data in place of the custom model

  Incorrect

  AutoML builds its own model from the data; it does not tune the team's custom training code.

• BAccelerators with more memory for each of the training runs

  Incorrect

  Bigger accelerators speed up each run, but someone still picks every configuration by hand.

• CA hyperparameter tuning job over the three training settings

  Correct

  Hyperparameter tuning runs many trials, uses the results of earlier trials, such as through Bayesian optimization, to choose the next values, and reports the best configuration.

• DA script that starts one training job for each combination

  Incorrect

  A fixed grid of jobs does not learn from earlier results, and the team still compares the runs itself.

Hyperparameters govern the training process itself. A hyperparameter tuning job on Agent Platform searches them across trials toward the target metric, which replaces trial and error by hand.