Free PMLE Practice Questions

Question 1

Unpredictable 45-second traffic bursts push an online prediction service beyond its 150 ms p95 latency objective. The Gemini Enterprise Agent Platform deployment keeps two replicas running and permits up to twelve. During a burst, the two replicas are compute-saturated; additional replicas need 75 seconds to load the model. Load tests with eight already-running replicas meet the objective, and quota is available. Without changing the model or the latency objective, which configuration change addresses the measured bottleneck?

Accepted Answer

Maintain eight ready replicas and allow scaling to twelve by raising the deployment's minimum replica count.

Answer

Raise the maximum replica count from twelve to twenty to provide a higher scaling ceiling.

Answer

Increase request timeouts so queued predictions can finish after additional replicas become ready.

Answer

Increase concurrency on each of the two replicas so they accept more requests during a burst.

Question 2

Promotion schedules for a product are fixed fourteen days in advance. Its daily demand depends on both recent demand and the planned discount. A SQL team keeps the sales history in BigQuery and wants a warehouse-native forecasting model that represents the time-series structure while incorporating those known future discounts. Which training and inference combination fits this requirement?

Accepted Answer

Train ARIMA_PLUS_XREG and call ML.FORECAST with future dates and the corresponding planned discount values.

Answer

Train ARIMA_PLUS and join planned discounts to the output of ML.FORECAST.

Answer

Train ARIMA_PLUS_XREG and call ML.PREDICT with future dates and discounts as ordinary prediction rows.

Answer

Train ARIMA_PLUS_XREG and call ML.FORECAST using historical discounts as substitutes for the future promotion schedule.

Question 3

For a checkout decision made at 09:00, an offline training query reconstructs the shopper's failed-payment count. Three feature records exist: count 2, event time 08:50, available to the prediction service at 08:52; count 4, event time 08:58, available at 09:04; count 5, event time 09:02, available at 09:03. The query currently chooses count 4 because its event time precedes checkout. How should the feature be reconstructed to retain useful history without giving training information that serving could not have known?

Accepted Answer

Use count 2, selecting the latest feature value that was actually available when the checkout decision was made.

Answer

Retain count 4 because the events occurred before checkout, even though their processed value arrived later.

Answer

Use count 5 because it is the most complete record available when the offline training dataset is assembled.

Answer

Remove failed-payment count from the entire training dataset whenever any event contributing to that feature can be processed late.

Question 4

Prompt revisions have stopped improving a Gemini support assistant. Error analysis shows that the correct, current manual passages are already retrieved, but the model repeatedly assigns the wrong resolution category under a stable company taxonomy. The team has thousands of expert-reviewed input-response examples and a separate evaluation set. Manuals change weekly; the categorization rules do not. What experiment targets the remaining failure while preserving access to current product facts?

Accepted Answer

Run supervised tuning on the reviewed examples, retaining retrieval for current manuals and evaluating against the baseline.

Answer

Retrain the document embedding model and enlarge retrieval results to supply more passages from each product manual.

Answer

Replace retrieval with periodic fine-tuning on newly issued manual text so product knowledge and task behavior are both encoded in model weights.

Answer

Keep the model unchanged and append complete manuals to each request instead of supplying the selected relevant passages.

Question 5

Yesterday's preprocessing output appears in today's training run even though new source records passed data-quality checks. The Gemini Enterprise Agent Platform Pipelines component receives the plain string source_uri='gs://ml-inputs/current.csv'; that object is overwritten daily. The component's image, parameters, and source_uri are unchanged, and its execution is marked as a cache hit. The team wants to retain reuse when inputs truly have not changed. What is the durable correction?

Accepted Answer

Publish immutable data snapshots and pass each snapshot's distinct identity as an explicit component input.

Answer

Change each run's display name but leave its component inputs unchanged.

Answer

Disable caching on the training component so it retrains even when preprocessing returns a previously cached artifact.

Answer

Disable caching throughout the pipeline so every component reruns, including those whose effective inputs are unchanged.

Question 6

A fraud team compares four score thresholds on the same fully labeled validation batch. Every flagged payment requires review, and the team can review at most 1,100 payments per batch. Its decision model assigns a $4 cost to each false positive and a $40 cost to each false negative; no other costs differ between thresholds. Results are:

Threshold 0.20: TP 380, FP 1,820, FN 20.
Threshold 0.40: TP 360, FP 640, FN 40.
Threshold 0.60: TP 320, FP 180, FN 80.
Threshold 0.80: TP 240, FP 60, FN 160.

Which threshold minimizes modeled error cost while respecting review capacity?

Accepted Answer

Threshold 0.60.

Answer

Threshold 0.20.

Answer

Threshold 0.40.

Answer

Threshold 0.80.

Question 7

An incident investigation confirms that instructions hidden in a retrieved vendor document caused a support agent to export customer records to an unauthorized external destination. Model Armor screens user prompts and final answers, but export-tool calls use a broadly privileged service account and do not check the requesting user's permissions. The export path remains enabled. Which response should take priority to contain this data-loss path?

Accepted Answer

Disable the export path immediately and require caller and destination authorization before restoring the tool.

Answer

Remove the malicious document, strengthen the system prompt, and allow the export workflow to continue.

Answer

Raise Model Armor sensitivity and retain export access while analysts investigate the resulting alerts.

Answer

Switch the model to a private endpoint while leaving the export integration and its permissions unchanged.

Question 8

OCR recognizes the text in a distributor's invoices accurately, yet the downstream extraction code frequently associates a line-item amount with the wrong description when suppliers change layouts. The team needs typed invoice fields and correctly grouped line items, has no custom-labeled training corpus, and wants the least custom development before a representative pilot. Which starting approach is most appropriate?

Accepted Answer

Pilot a pretrained Document AI invoice processor, checking its extracted fields and line-item groups against representative invoices.

Answer

Keep Cloud Vision OCR and build supplier-specific rules that infer line-item associations from text coordinates.

Answer

Build a labeled invoice corpus and train a custom Document AI extractor to cover the supplier layouts.

Answer

Train an image object-detection model to locate invoice cells, then develop custom logic to reconstruct each row.

Question 9

A profiler from a custom image-training job shows long gaps between GPU kernels and GPU utilization near 20%. One decoding worker saturates a CPU core while other CPU cores are idle; storage reads are fast. Replacing the real input stream with pre-generated tensors raises GPU utilization above 90% without changing the model or batch size. Given this controlled comparison, where should the engineer spend the next optimization effort?

Accepted Answer

Parallelize decoding and prefetch prepared batches to keep the accelerator supplied with data.

Answer

Move to a GPU with greater tensor throughput while retaining the existing decoding and input-loading implementation.

Answer

Enable mixed-precision arithmetic in the model so the currently active GPU kernels complete in less time.

Answer

Add data-parallel GPU workers while retaining the same CPU decoding capacity and shared input-loading path.

Question 10

A recommendation team proposes a 50/50 endpoint traffic split to compare weekly purchase conversion per user. Individual users make many requests and can receive recommendations from both model versions within the same shopping journey. Serving reliability has already been validated; the remaining question is whether the candidate changes user behavior. Which assignment design supports that comparison?

Accepted Answer

Randomly assign each user to one model and preserve that assignment for the whole experiment.

Answer

Split each user's requests independently; attribute purchases to the model serving checkout.

Answer

Shadow every request to the candidate and compare purchases while users continue receiving the incumbent's recommendations.

Answer

Serve the incumbent to all users for one week and the candidate to all users the following week.

Free PMLE Practice Questions

Domain 1: Architecting low-code AI solutions (~13%)

Question 1

Question 2

Domain 2: Collaborating within and across teams to manage data and models (~16%)

Question 3

Question 4

Domain 3: Scaling prototypes into ML models (~21%)

Question 5

Question 6

Domain 4: Serving and scaling models (~20%)

Question 7

Question 8

Domain 5: Automating and orchestrating ML pipelines (~18%)

Question 9

Domain 6: Monitoring AI solutions (~13%)

Question 10

That's 10 of 1,030