Free PMLE flashcards

What model and Google Cloud product fit nightly univariate forecasts for thousands of series owned by a SQL-only team?

ARIMA models univariate temporal autocorrelation directly, and BigQuery ML lets a SQL-fluent team train and forecast many per-entity series in-place at low cost. Results stay in BigQuery, so dashboards read them without any Python serving pipeline or data movement.

Which approach gives high accuracy, low build effort, and faithful per-decision feature attributions on structured data?

AutoML tabular automates model selection and tuning for strong accuracy with little coding, and its built-in feature attributions quantify how much each input moved an individual prediction. That measurable per-decision contribution is what regulated interpretability demands.

Common misconceptionThat an LLM's fluent explanation counts as a genuine feature attribution. A generated narrative can state reasons the scoring model never actually used.

Why does a GPU-backed endpoint fail to scale out when GPUs are saturated but the default autoscaling signal tracks CPU?

An autoscaler only adds replicas when its chosen metric crosses the target. GPU inference is bound by GPU duty cycle, not CPU, so on a GPU-bound workload the CPU stays low and the CPU-based trigger never fires even as p99 latency climbs. Switching the scaling signal to the GPU duty-cycle metric ties scale-out to the resource that actually saturates.

How do you tune Vertex AI endpoint autoscaling to absorb a sharp morning spike without paying for idle replicas overnight?

Keep a small minReplicaCount above zero so a few warm replicas answer the first burst with no cold-start delay, raise maxReplicaCount well above peak so the endpoint can scale out for the spike, and lower the target utilisation so the autoscaler adds replicas before the request queue saturates. Together these cover the spike economically without holding peak capacity through the quiet hours.

Common misconceptionThat adding client-side retries fixes request timeouts. Retries provision no extra capacity and can add load during an already-saturating scale-out.

What mechanism automatically starts a Cloud Build pipeline when code is merged to a Git branch?

A Cloud Build trigger bound to the repository branch. It subscribes to repository push and merge events on the chosen branch and invokes the build configuration automatically, with no manual step. This removes the latency and missed-step risk of polling or hand-run commands.

When model decay arrives unpredictably rather than on a calendar, what should trigger retraining?

Tie retraining to a monitored signal: fire it when prediction performance or input drift crosses a defined threshold. Because unpredictable concept drift cannot be anticipated by a clock, a threshold trigger fires exactly when the model genuinely degrades, conserving compute while keeping the model fresh only when accuracy actually falls.

Common misconceptionThat retraining every night guarantees a fresh model. It burns compute on runs where nothing has actually degraded.

What engine fits a one-off, multi-terabyte join and window-aggregation backfill over Parquet in Cloud Storage when the team wants standard SQL and no clusters?

BigQuery. It is a serverless analytics engine that scales terabyte joins and window functions automatically, reads Parquet through external tables, and needs no provisioned cluster. Loading the Parquet as an external table lets a single SQL query join it against an existing dimension table, satisfying scale, SQL, and no-infrastructure constraints together.

How can free-form generative outputs be scored automatically for relevance and faithfulness when single-label accuracy does not apply?

Use an LLM-as-a-judge to score each output against defined qualitative criteria such as relevance and faithfulness. Generative text has many acceptable forms and no single ground-truth label, so label-based metrics cannot capture quality. An LLM judge gives repeatable, automated ratings that scale across a dataset.

Common misconceptionThat a lower training perplexity proves generated summaries are relevant and faithful. Perplexity measures token-prediction fit, not whether the output stays on topic or true to the source.