GCP-PMLE Sample Questions & Answers

A Cloud Build trigger connected to your Git repository (for example, GitHub) starts a build only when new code is pushed, so retraining runs automatically (no manual intervention), only when the architecture actually changes (minimal compute cost), and the code is version-controlled in Git. Submitting jobs manually with gcloud requires intervention, Cloud Storage is not version control, and a daily Composer sensor adds cost and polling overhead.

In an Agent Platform Workbench instance the BigQuery client library is preinstalled, and the %%bigquery magic runs the SQL query, converts the results to a pandas DataFrame and can save it to a variable in one step (using the BigQuery Storage API for fast downloads). Exporting to CSV via Drive, local download or Cloud Storage adds unnecessary manual steps.

Google Cloud Pipeline Components provide a prebuilt BigqueryQueryJobOp that launches a BigQuery query job, waits for it to finish and exposes the result as a pipeline artifact for downstream steps. Reusing a prebuilt component is the easiest option; writing your own script or custom component adds code to build and maintain, and running the query manually in the console is not part of the automated pipeline.

Temperature readings are a time series. Readings an hour apart are highly correlated, so a random split puts near-copies of the test-period data into the training data. This data leakage made the test accuracy (97%) look much better than the real accuracy in production (66%). Google's guidance is to use a chronological split for time-series tasks. It also says to compute transformation statistics (for normalization or bucketizing) from the training split only and apply them to the validation and test splits. Applying transformations before splitting is itself a source of leakage. Adding more test data, or normalizing the training and test data as separate steps, doesn't remove the leakage that the random split causes.

A ResourceExhaustedError (OOM) means the tensors for one training step do not fit in GPU memory. Activation memory scales with the batch size, so reducing the batch size is the standard fix that keeps the model and input resolution unchanged. Changing the optimizer or learning rate does not materially change memory use, and reducing the image shape would change the model input and can hurt accuracy.

Precision = TP / (TP + FP). To increase precision you raise the classification threshold so that fewer borderline images are labeled "car", which reduces false positives. The trade-off is that more true cars fall below the threshold, so false negatives increase and recall decreases. Increasing recall, increasing false positives or decreasing false negatives all move in the opposite direction.

Firebase Cloud Messaging (FCM) is designed to send notifications to individual users and devices at very large scale. Register each user with FCM, and send that user a message when the model predicts that their own balance will drop below $25. A Pub/Sub topic for each user doesn't scale to millions of customers. A project can have at most 10,000 topics, and you can't request an increase to this limit. This is true whether a Cloud Run function or an App Engine app sends the notifications. Sending a notification when the average of all users' predictions drops below $25 ignores each user's own forecast.

When numeric features have very different ranges, gradient descent takes uneven steps and struggles to converge. Normalization (for example, z-score or scaling to a common range) puts features on similar scales so gradient optimization can move the weights to a good solution. Combining features, removing features with missing values or changing the split do not address the scale mismatch.

A Gemini model on Gemini Enterprise Agent Platform can classify raw transcript text by product with prompting, and supervised tuning on your labeled transcripts improves accuracy when needed, with no feature engineering or model code. (This replaces the retired AutoML Natural Language / AutoML text products.) The Cloud Natural Language API cannot be trained on your product categories, and building a custom training job or a keyword-extraction pipeline requires far more preprocessing and development time.

Transfer learning starts from an established, pre-trained TensorFlow text classification model and fine-tunes it on your own labeled support requests, which saves time while keeping full control of the TensorFlow code, serving and deployment on managed custom training. Using the model as-is ignores your classes, and the Natural Language API or a tuned Gemini model does not give you control of the model code.