Read this the night before or on exam morning. It assumes you’ve studied the notes. It’s here to refresh, not to teach.
Exam day
- 50 to 60 multiple choice and multiple select questions in 2 hours. Fee $200. English and Japanese. Online or test-centre proctored. No prerequisites; 3+ years of industry experience, 1+ on Google Cloud is recommended.
- Google reports pass or fail rather than a published passing score.
- Official section weights: low-code AI solutions 13%, collaborating on data and models about 14%, scaling prototypes into models about 18%, serving and scaling about 20%, automating and orchestrating pipelines about 22%, monitoring about 13%.
- About 2 minutes per question. Flag long scenarios and return.
- The current guide includes generative AI: Model Garden, Agent Builder, RAG, fine-tuning and evaluating generative output.
- Naming: Google is rebranding Vertex AI as the Gemini Enterprise Agent Platform. The exam guide still says Vertex AI.
How to read a PMLE question
| The question says | Pick |
|---|---|
| Minimal effort, no ML expertise, fastest | Pre-trained API, then BigQuery ML or AutoML, before custom training |
| Data is already in BigQuery, SQL skills | BigQuery ML |
| Own labelled images, text, video or tabular | AutoML |
| Answers from our documents, up to date, citations | RAG with Vertex AI Search or RAG Engine |
| Change the model’s style or output format | Fine-tuning (SFT or LoRA) |
| Must explain predictions | Explainable AI, or a simpler interpretable model |
| Predictions can be computed overnight | Batch prediction |
| Millisecond response | Online endpoint with warm replicas and Feature Store lookup |
| Serving data differs from training data | Skew monitoring, fix preprocessing |
| Inputs change over time | Drift monitoring, retrain |
| Repeatable, scheduled, auditable | Vertex AI Pipelines, ML Metadata |
| Keep traffic private | Private Service Connect endpoint, VPC Service Controls |
| Reduce cost | Spot VMs for training (with checkpoints), right-size, batch instead of online, quantise or distil |
Elimination tricks
- Custom training is the wrong answer when an API, BQML or AutoML meets the need.
- Avoid answers that export data out of BigQuery without reason.
- Raising accuracy by adding compute rarely fixes a data or leakage problem.
- Retraining does not fix training-serving skew caused by a pipeline bug.
- Evaluate on a held-out set that matches production. Chronological split for time series.
- In choose-two questions, each pick should cover a different part of the solution.
BigQuery ML model types
| Problem | model_type |
|---|---|
| Number | LINEAR_REG, BOOSTED_TREE_REGRESSOR |
| Binary or multiclass | LOGISTIC_REG, BOOSTED_TREE_CLASSIFIER, DNN_CLASSIFIER |
| Segments | KMEANS |
| Recommendations | MATRIX_FACTORIZATION |
| Time series | ARIMA_PLUS (XREG for covariates), TimesFM |
| Anomaly | AUTOENCODER, ML.DETECT_ANOMALIES |
| Dimension reduction | PCA |
| LLM over rows | Remote model, ML.GENERATE_TEXT, ML.GENERATE_EMBEDDING |
TRANSFORM keeps preprocessing in the model. ML.PREDICT is batch. Online serving means Model Registry plus an endpoint.
Which tool
| Need | Choose |
|---|---|
| Streaming or large preprocessing | Dataflow (Apache Beam) |
| Tabular preprocessing | BigQuery SQL |
| Preprocessing shared with serving | TensorFlow Transform, TFX, Feature Store |
| Shared, reusable features | Vertex AI Feature Store (BigQuery-based) |
| Find and mask PII | Sensitive Data Protection (Cloud DLP), policy tags |
| Governed notebooks | Vertex AI Workbench instances. Serverless: Colab Enterprise |
| Compare runs | Vertex AI Experiments. Training curves: TensorBoard |
| Lineage and audit | Vertex ML Metadata, Model Registry |
| ML workflow | Vertex AI Pipelines (KFP or TFX) |
| ML as part of a bigger data DAG | Cloud Composer |
| CI for ML code | Cloud Build |
| Self-hosted MLflow | Cloud Run or GKE, Cloud SQL, Cloud Storage |
| Many-GPU gradient sync | Reduction Server, Horovod, NCCL |
Training facts
- Worker pool 0 is the chief, 1 workers, 2 parameter servers, 3 evaluators.
- Data parallelism by default. Model parallelism only if the model doesn’t fit.
- Spot VMs need checkpoints. The job’s service account (not yours) needs data access.
- Hyperparameter tuning is Bayesian by default. Use log scale for learning rate. More trials beat more parallelism. Tune on validation, not test.
- NaN loss: learning rate or bad data. Big train-validation gap: overfit. Perfect score: leakage. OOM: smaller batch, mixed precision or bigger device.
- CPU for classical ML, GPU for general deep learning, TPU for large matrix-heavy training, edge (TF Lite) for devices.
Serving facts
- Endpoints hold multiple deployed models with a traffic split summing to 100%. Canary, A/B and rollback all use it. Shadow testing sends no user-visible output.
- Private access: Private Service Connect (current), VPC peering (older). Dedicated public endpoints give isolation and performance.
- Autoscaling targets utilisation (CPU default 60%, GPU duty cycle). Keep min replicas at 2 or more for availability.
- Batch: Vertex batch prediction, BigQuery ML, Dataflow RunInference (streaming), Dataproc.
- Optimise: quantisation, pruning, distillation, compilation (TensorRT, XLA, ONNX), batching, caching.
- Errors: 429 capacity or quota, 5xx container or model, 400 schema mismatch.
Monitoring facts
| Check | Compares | Fix |
|---|---|---|
| Training-serving skew | Serving data with training data | Fix preprocessing or the data source |
| Prediction (input) drift | Serving now with earlier serving | Retrain on recent data |
| Feature attribution drift | Feature importance now with the baseline | Investigate and retrain |
- Jensen-Shannon divergence for numeric features, L-infinity for categorical. Thresholds per feature, sampling rate and interval are configurable. Alerts go to email, Logging, Monitoring and Pub/Sub.
- Labels arrive late, so use skew and drift as leading indicators, then confirm with ground truth by joining logged predictions in BigQuery.
- Explainable AI: Sampled Shapley (any model), Integrated Gradients (differentiable), XRAI (images). The baseline matters.
- Fairness: slice metrics. Parity, equal opportunity and equalised odds can conflict, so choose with stakeholders. Model cards document limits.
- LLM risks: prompt injection, leakage, jailbreaks. Use Model Armor, safety filters, grounding, least-privilege tools.
Retraining triggers
Schedule, new data arrival, performance decay (needs labels), drift or skew alert, code change, manual approval. Always gate: validate data, compare to the champion, then canary.