Try 10 Real Questions From Google Cloud Professional Machine Learning Engineer
These are unedited questions from the full 700-question set, pulled from across every exam domain. Click a question to reveal the answer and the full rationale for every option, exactly as it appears in the PDF.
Five years of sales history, promotion flags, and holiday indicators for 500 grocery stores already sit in BigQuery. Daily units must be forecast 30 days ahead, and the two-person analytics team has no infrastructure budget. Which approach delivers reliable forecasts with the least operational overhead?
Show answer & full rationale
A: Exporting 500 series and managing a container adds engineering and infrastructure work that BigQuery ML removes entirely.
B: An LSTM demands custom code, tuning, and cluster management that a two-person team cannot sustain for a standard seasonal forecast.
C: AutoML forecasting is capable but costs more operational overhead than an in-warehouse ARIMA_PLUS model for this workload.
D (Correct): ARIMA_PLUS trains one model per store from a single SQL statement, handles seasonality and holiday effects natively, and never moves the data out of BigQuery.
Three data types must coexist for one project: 40 TB of transactional tables, 6 million product description documents, and 12 million catalogue photographs. Training, experimentation, and serving all need governed access. How should this be organised?
Show answer & full rationale
A (Correct): This puts each data type in its natural store while object tables give BigQuery governed metadata access to the Cloud Storage files, so IAM and lineage stay unified.
B: Inlining binary media into a warehouse inflates storage cost and gives no benefit over object references.
C: This abandons the query, partitioning, and column-level controls the tabular workload depends on.
D: Managed datasets are a training convenience, not a system of record.
Clickstream events arrive continuously and features must reflect the last five minutes of user activity. Which preprocessing tool is the right fit?
Show answer & full rationale
A: Micro-batching in SQL introduces latency and duplicates work at each boundary.
B (Correct): Beam on Dataflow is built for unbounded data with event-time windowing, which is what a rolling five-minute feature requires.
C: This is the same micro-batch compromise with a cluster attached.
D: This neither scales nor handles event-time semantics.
Energy demand must be predicted hourly for the next seven days. Five years of hourly history exist with daily and annual cycles, and planned maintenance outages are known in advance. Which model choice is best justified?
Show answer & full rationale
A: Language models are not calibrated numeric forecasters and offer no prediction intervals.
B (Correct): The outages are known ahead of time, which is precisely the situation exogenous regressors exist for, and a classical seasonal model handles the strong cyclical structure well.
C: This captures the seasonality but cannot use the maintenance schedule.
D: This can work but discards the decomposition and interval estimates a seasonal model gives directly.
Regulators require that the reasoning behind any declined insurance policy be reconstructable from the model itself, and applicants must receive a written reason. What modelling choice best satisfies this?
Show answer & full rationale
A (Correct): When the explanation must come from the model rather than from a post-hoc approximation, an inherently additive structure is what makes reconstruction possible.
B: Post-hoc attributions approximate the model and are harder to defend as the reasoning itself.
C: Attention weights are not a faithful account of the decision.
D: Averaging across families makes any single explanation unavailable.
Risk scores for 60 million accounts are recomputed once a week and written to a table that downstream systems read. No interactive consumer exists. Which serving approach is correct?
Show answer & full rationale
A: This maintains a serving stack and adds per-request overhead for a workload with no latency requirement.
B: Sixty million individual invocations is far more expensive and slower than one batch job.
C: This achieves the same result with substantially more operational work.
D (Correct): Batch inference matches the access pattern exactly and avoids paying for idle serving capacity between weekly runs.
Tokenisation before inference and a business-rule adjustment afterwards must run identically for every caller. Where should this logic live?
Show answer & full rationale
A: Every caller then implements it separately, which is how serving skew begins.
B: This works but adds two network hops and a second component to version.
C (Correct): Binding the transformations to the model artifact guarantees every caller gets the same behaviour and that the logic is versioned alongside the weights.
D: A warehouse view cannot participate in an online request path.
A pipeline must halt before deployment whenever the candidate model fails to beat the incumbent on a held-out set. How should this be expressed?
Show answer & full rationale
A: This deploys the worse model first and relies on someone reacting.
B: Manual gates do not scale and are inconsistently applied.
C: Waiting for production metrics exposes users to a known-inferior model.
D (Correct): Encoding the gate as a conditional inside the pipeline makes the promotion rule executable and auditable rather than a manual convention.
A workflow must coordinate a BigQuery export, an on-premises file transfer, a training job, and a Salesforce update. What orchestration choice fits?
Show answer & full rationale
A (Correct): Airflow's strength is cross-system orchestration with a large operator ecosystem, which is exactly what a workflow touching four different systems needs.
B: ML pipelines excel at containerised ML steps with artifact lineage rather than at broad cross-system integration.
C: This builds bespoke orchestration with no shared retry, backfill, or visibility model.
D: Fixed offsets are a timing assumption that eventually breaks.
Credentials on a training job would permit writes almost anywhere. What most directly prevents exfiltration to a bucket outside the organisation?
Show answer & full rationale
A: IAM controls what the identity may read but does not restrict where it may write outside the organisation.
B: The workload decrypts legitimately and can then write the plaintext anywhere.
C (Correct): A service perimeter restricts which resources the workload can reach, which is the control that actually blocks the exfiltration path.
D: Logging records the exfiltration rather than preventing it.
Google Cloud Professional Machine Learning Engineer — $19.99, instant PDF download with full rationale on every answer.
View the practice questions →