PMLE — Google Cloud Professional Machine Learning Engineer Cheat Sheet

Compact PMLE Cheat sheet for Google Cloud machine learning engineering: Vertex AI, data pipelines, MLOps, security, monitoring, and decision points.

Use the tables for a quick pre-exam check. Expand a topic’s notes for explanations, examples, and additional distinctions.

Exam-use mental model

This independent Cheat Sheet supports candidates preparing for Google Cloud Professional Machine Learning Engineer - 2026 Guide from Google Cloud, exam code PMLE. Use it to rehearse service selection, ML lifecycle decisions, and common exam traps.

For most PMLE scenarios, choose the answer that best satisfies:

  1. Business objective first: metric, latency, freshness, cost, risk, explainability.
  2. Managed Google Cloud service when practical: reduce custom infrastructure unless requirements demand it.
  3. Reproducible ML lifecycle: version data, code, parameters, artifacts, metrics, and deployments.
  4. Secure by design: least privilege, private data paths, encryption, auditability.
  5. Monitor after deployment: prediction quality, drift, skew, latency, errors, and retraining triggers.
    flowchart LR
	    A[Problem framing] --> B[Data ingestion and validation]
	    B --> C[Feature engineering]
	    C --> D[Training or tuning]
	    D --> E[Evaluation and model registry]
	    E --> F[Deployment: online, batch, or embedded]
	    F --> G[Monitoring and explainability]
	    G --> H[Retraining pipeline]
	    H --> C
Notes and examples

The PMLE mental model

A strong PMLE answer usually follows the full ML lifecycle, not just model training.

Lifecycle stageWhat to decideHigh-yield PMLE focus
Business framingWhat outcome matters?Translate business goals into measurable ML objectives and constraints.
Data sourcingWhat data is available and trustworthy?Use appropriate storage, pipelines, schemas, labels, and access controls.
Feature preparationHow will inputs be transformed?Prevent leakage, handle missing values, manage categorical features, and preserve training-serving consistency.
Model selectionWhat approach is practical?Choose AutoML, BigQuery ML, custom training, pretrained APIs, or foundation models based on requirements.
Training and tuningHow will the model improve?Use correct split strategy, metrics, hyperparameter tuning, distributed training, and regularization.
EvaluationIs the model good enough?Validate on holdout data, slices, fairness dimensions, latency, cost, and business impact.
DeploymentHow will predictions be served?Choose batch, online, streaming, endpoint, container, or custom serving patterns.
MonitoringHow will issues be detected?Track skew, drift, prediction quality, latency, errors, and retraining triggers.
GovernanceIs it secure and responsible?Use IAM, encryption, auditability, privacy controls, explainability, and human review where needed.

Fast decision rule

When a question gives you multiple technically possible answers, prefer the one that is:

  1. Managed when requirements are standard.
  2. Reproducible when training or deployment must be repeatable.
  3. Least privilege when security is involved.
  4. Observable when production reliability matters.
  5. Cost-aware when scale, idle resources, or accelerators are mentioned.
  6. Aligned to the metric when model quality is the issue.

High-yield Google Cloud service selection

Core ML and MLOps services

RequirementPreferUse whenAvoid when
End-to-end managed ML lifecycleVertex AINeed managed training, tuning, model registry, endpoints, pipelines, metadata, monitoringYou only need simple SQL analytics or non-ML batch processing
Low-code model buildingVertex AI AutoMLTabular, image, text, or similar tasks with limited custom architecture needsNeed full model architecture control, custom loss, specialized training loop
Custom model trainingVertex AI custom trainingNeed custom code, framework, containers, distributed training, GPUs/TPUsSQL-native ML or AutoML meets requirements
SQL-native ML in warehouseBigQuery MLData already in BigQuery; analysts use SQL; batch scoring is acceptableComplex custom deep learning pipeline or low-latency online serving is required
Pretrained perception or language APIGoogle Cloud AI APIs / Vertex AI foundation modelsNeed fast integration without training from scratchNeed domain-specific model behavior, private fine-tuning, or custom serving logic
Pipeline orchestration for MLVertex AI PipelinesNeed reproducible ML steps, artifacts, metadata, lineage, scheduled retrainingGeneral non-ML workflow is primary concern
General workflow orchestrationCloud Composer or WorkflowsNeed broad DAG orchestration across servicesNeed ML-native lineage, artifacts, experiments, and model metadata
Experiment trackingVertex AI Experiments / TensorBoardNeed compare runs, parameters, metrics, artifactsOne-off notebook work with no reproducibility requirement
Model catalog and promotionVertex AI Model RegistryNeed approve, version, deploy, rollback modelsModel is never reused or governed
Online predictionVertex AI endpointsNeed managed low-latency serving, scaling, traffic splittingScoring can be delayed or done in bulk
Batch predictionVertex AI batch prediction or BigQuery ML batch scoringNeed score large datasets asynchronouslyInteractive user request needs immediate response
Containerized custom inferenceVertex AI custom prediction container, Cloud Run, or GKENeed custom pre/post-processing or nonstandard serving stackStandard Vertex AI serving container is enough
Notes and examples

Data and analytics services for ML

RequirementPreferWhy it fits
Large analytical warehouse, SQL features, BI, BQMLBigQueryServerless analytics, feature extraction, training data assembly, batch scoring
Object storage for raw data, images, model artifactsCloud StorageDurable storage for datasets, exports, checkpoints, training packages
Stream ingestionPub/SubDecouples producers and consumers; supports event-driven pipelines
Stream/batch transformationsDataflowManaged Apache Beam for scalable ETL, windowing, streaming features
Existing Spark/Hadoop jobsDataprocManaged Spark/Hadoop when reusing ecosystem code
Metadata, governance, discoveryDataplex / Data Catalog capabilitiesHelps classify, discover, and govern data assets
SecretsSecret ManagerAvoid hardcoded credentials in notebooks, containers, or pipelines
Container imagesArtifact RegistryVersion custom training and serving containers
CI/CDCloud Build plus deployment toolingBuild, test, and promote ML code and containers

Google Cloud service map for PMLE

NeedCommon Google Cloud fitWatch for
End-to-end ML platformVertex AITraining, pipelines, model registry, endpoints, batch prediction, experiments, monitoring.
SQL-based analytics and modelingBigQuery and BigQuery MLGood for large structured data already in BigQuery; not always best for complex custom deep learning.
Object storage for datasets and artifactsCloud StorageRaw files, images, exports, model artifacts, staging data.
Batch and streaming data processingDataflowApache Beam pipelines, scalable ETL, streaming feature generation.
Spark or Hadoop workloadsDataprocExisting Spark jobs, migration of Hadoop/Spark pipelines, large-scale transformations.
Event ingestionPub/SubDecoupled streaming ingestion, event-driven ML pipelines.
Workflow orchestrationVertex AI Pipelines, Cloud Composer, WorkflowsChoose based on ML-native pipeline needs versus general orchestration.
Container build and artifact storageCloud Build and Artifact RegistryCI/CD, reproducible containers, secure image management.
Custom servingVertex AI endpoints, GKE, Cloud RunVertex AI for managed prediction; GKE/Cloud Run for custom app-level requirements.
Monitoring and logsCloud Monitoring and Cloud LoggingLatency, error rates, resource metrics, pipeline failures, service health.
Secrets and keysSecret Manager and Cloud KMSAvoid secrets in code, notebooks, containers, or environment files.
Identity and accessIAM and service accountsLeast privilege, separation of duties, workload-specific permissions.
Data protectionSensitive Data Protection, VPC Service Controls, CMEK where requiredUse when data sensitivity, boundaries, or encryption control are explicit requirements.

PMLE decision tables

AutoML vs custom training vs BigQuery ML vs pretrained model

Scenario clueBest fitReasoning
“Fast baseline,” “limited ML expertise,” “standard tabular/image/text problem”Vertex AI AutoMLManaged feature processing, training, tuning, and evaluation
“Custom loss,” “custom neural network,” “special preprocessing,” “research model”Vertex AI custom trainingFull code and framework control
“Data is in BigQuery,” “team uses SQL,” “batch predictions,” “no custom serving”BigQuery MLKeeps ML close to warehouse data and SQL workflows
“Need sentiment/OCR/speech/translation quickly”Pretrained Google Cloud AI APIsNo training pipeline required
“Need private enterprise answers from documents”Vertex AI foundation model with grounding/RAGInjects current/private facts without training model from scratch
“Need specialized language/style/task adaptation”Tuning on Vertex AI, when supportedChanges behavior more than prompting, less work than full custom training
“Strict control over weights, architecture, training data, serving”Custom model on Vertex AIRequired when managed abstractions are insufficient
Notes and examples

Online vs batch prediction

RequirementChooseWatch for
User-facing request/responseOnline prediction endpointLatency, autoscaling, model size, input validation
Millions of rows scored overnightBatch predictionThroughput, output location, idempotency
Scores joined with warehouse tablesBigQuery ML prediction or batch output to BigQuerySQL governance and reproducibility
Event-driven near-real-time scoringPub/Sub + Dataflow + online endpoint or streaming architectureBackpressure, retry behavior, duplicate handling
Very low latency with custom serving logicCloud Run or GKE may be consideredMore operational responsibility than managed endpoint
Model embedded on device or edgeExported model format if supportedUpdate strategy, device constraints, monitoring limitations

Pipeline orchestration choice

NeedPreferExam distinction
ML steps with artifacts, metadata, lineageVertex AI PipelinesPMLE default for reproducible MLOps
Airflow DAG already orchestrates enterprise data platformCloud ComposerGood for heterogeneous scheduled workflows
Simple service-to-service workflowWorkflowsLightweight orchestration, not ML-specific
Pure ETL transform at scaleDataflowData processing engine, not experiment tracker
CI/CD build-test-deployCloud BuildBuild automation, not training lineage by itself

Data preparation and feature engineering

Data split patterns

Data typeRecommended splitCommon trap
Independent tabular rowsRandom or hash-based splitNon-reproducible random split causing changing metrics
Time seriesTime-based split: train on past, validate on futureRandom split leaks future information
User behaviorSplit by user/entity when leakage across rows is possibleSame user appears in train and test
Image/text duplicatesDeduplicate or group before splitNear-duplicates inflate evaluation
Imbalanced classificationStratified split when appropriateMinority class disappears from validation/test
Streaming dataHold out later time windowsOffline test set does not match production freshness
Notes and examples

Stable BigQuery split pattern:

SELECT
  *,
  CASE
    WHEN MOD(ABS(FARM_FINGERPRINT(CAST(customer_id AS STRING))), 10) < 8 THEN 'TRAIN'
    WHEN MOD(ABS(FARM_FINGERPRINT(CAST(customer_id AS STRING))), 10) = 8 THEN 'VALIDATE'
    ELSE 'TEST'
  END AS split
FROM `project.dataset.source_table`;

Feature transformation location

Transform locationUse whenRisk
BigQuery SQLBatch features, warehouse-native joins, aggregationsTraining-serving skew if online path reimplements logic differently
Dataflow / Apache BeamStreaming features, large-scale ETL, unified batch/stream processingOperational complexity if simple SQL is enough
tf.Transform-style pipeline stepNeed identical training and serving transformsMore pipeline complexity
Model preprocessing layerTransform must be packaged with modelCan increase serving latency
Feature store / online feature servingNeed consistent offline/online features and low-latency lookupRequires governance around freshness and keys
Application codeSimple request formattingHigh skew risk if business logic diverges from training

Data quality checks to rehearse

CheckWhy it matters
Schema validationDetects missing, renamed, or type-changed fields
Range checksFinds impossible values, unit errors, and outliers
Null/missingness trackingMissingness may be predictive or indicate broken ingestion
Label validationIncorrect labels can cap model performance
Class distributionPrevents misleading accuracy on imbalanced data
Train/serve feature parityReduces skew between offline training and online prediction
Duplicate detectionPrevents leakage and inflated metrics
Time-window correctnessPrevents future data from entering features
PII/sensitive data classificationSupports least privilege and privacy controls

Data preparation and feature engineering

Good PMLE answers protect model quality before training begins.

TopicReview pointsCandidate mistakes
Data qualityValidate schema, ranges, missingness, duplicates, outliers, label consistency.Training on unvalidated data because the model “can learn around it.”
Data splitsUse train/validation/test; time-based splits for time-dependent data; group splits for related records.Random split when users, devices, households, or future events leak across splits.
Label qualityCheck labeling instructions, consensus, inter-rater agreement, delay between event and label.Treating noisy labels as ground truth without validation.
Feature leakageExclude fields unavailable at prediction time or derived from the target.Including post-event data, future aggregates, or target-encoded features incorrectly.
Missing valuesImpute consistently; add missingness indicators when meaningful.Using different missing-value logic in training and serving.
Categorical featuresUse one-hot, embeddings, hashing, or native handling depending on model type and cardinality.One-hot encoding extremely high-cardinality features without considering memory or generalization.
Numerical featuresScale when using distance-based models, linear models, neural networks, or gradient-sensitive methods.Scaling unnecessarily for tree models, or fitting scalers on all data before splitting.
Text/image/audioUse appropriate preprocessing, pretrained models, embeddings, or specialized architectures.Building custom models when pretrained APIs or foundation models would meet requirements.
Feature reuseCentralize transformations and feature definitions where possible.Duplicating feature logic across training and serving code.

Training-serving skew

Training-serving skew occurs when the model sees one feature distribution or transformation during training and a different one during prediction.

Common causes:

  • Different preprocessing code paths for training and serving.
  • Time-window aggregations computed differently offline and online.
  • Missing values handled differently in production.
  • Categorical vocabularies not frozen or versioned.
  • Feature values available in batch training but unavailable at request time.
  • Data schema changes not detected before prediction.

Best review answer: use shared transformation logic, versioned artifacts, schema validation, pipeline automation, and monitoring for skew or drift.

Modeling and evaluation reference

Metric selection

Problem typePrefer metricsUse whenTrap
Balanced classificationAccuracy, log loss, AUCClasses are roughly balanced and error costs similarAccuracy hides minority-class failure
Imbalanced classificationPrecision, recall, F1, PR AUCFraud, churn, abuse, rare disease, anomaly review queuesROC AUC can look good while precision is poor
Ranking/recommendationNDCG, MAP, precision@k, recall@kTop results matter more than all predictionsOptimizing overall accuracy instead of ranked utility
RegressionRMSE, MAE, R-squaredPredict continuous valuesRMSE over-penalizes large errors; MAE may hide severe outliers
ForecastingMAE, RMSE, MAPE, weighted errorsTime-dependent demand or capacityRandom split and MAPE issues near zero values
ClusteringSilhouette, Davies-Bouldin, business validationNo labels availableTreating unsupervised score as proof of business value
Generative AIGroundedness, factuality, safety, task success, human preferenceOpen-ended outputsEvaluating only fluency and ignoring hallucination
Notes and examples

For binary classification:

\[ \text{Precision}=\frac{TP}{TP+FP} \]\[ \text{Recall}=\frac{TP}{TP+FN} \]\[ F1=2\cdot\frac{\text{Precision}\cdot\text{Recall}}{\text{Precision}+\text{Recall}} \]\[ \text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN} \]

Threshold and error-cost decisions

RequirementDecision
False positives are expensiveIncrease precision; raise threshold
False negatives are expensiveIncrease recall; lower threshold
Human review capacity is limitedOptimize precision@k or top-k workload
Regulatory or customer-impacting decisionFavor explainability, monitoring, audit logs, and human review
Need calibrated probabilitiesEvaluate calibration, not just ranking
Class distribution shifts in productionMonitor prediction distribution and labels when available

Model family shortcuts

Model familyStrengthWeaknessExam clue
Linear/logistic regressionSimple, interpretable, fastLimited nonlinear patternsBaseline, explainability
Tree-based models / boosted treesStrong for tabular data, handles nonlinearitiesLess ideal for raw images/textStructured business data
Deep neural networksFlexible, high capacityNeeds more data/tuningImages, text, complex signals
CNNsSpatial patternsImage-specific architectureVision workloads
TransformersText, multimodal, generative tasksCost, latency, safety evaluationNLP, LLM, embeddings
Matrix factorization / two-tower retrievalRecommendations and retrievalCold-start handling neededUsers/items, candidate generation
Time-series modelsTemporal structureMust respect time orderingForecasting demand, capacity, traffic

Generative AI and foundation-model decisions

RequirementPreferWhy
Prototype text generation, summarization, extractionVertex AI foundation model promptingFastest path; no training data required
Need answers grounded in private documentsRetrieval-augmented generationKeeps facts in external corpus and reduces stale knowledge risk
Need enterprise search over documentsVertex AI Search / grounding-oriented architectureManaged retrieval and relevance features
Need domain-specific style or task behaviorModel tuning, if supported for selected modelAdjusts behavior beyond prompt engineering
Need deterministic structured extractionPrompt with schema, validation, and post-processingLLM output should still be validated
Need safety controlsSafety settings, content filters, allowlists, human reviewDo not rely only on prompt wording
Need evaluate generated answersHuman evaluation plus automated checksFluency is not enough
Notes and examples

Common generative AI traps:

  • Using fine-tuning to add frequently changing facts when RAG is more appropriate.
  • Ignoring grounding, citation, and hallucination checks.
  • Sending sensitive data to prompts without access control and logging review.
  • Evaluating only on “good looking” outputs instead of task-specific test sets.
  • Forgetting latency and token cost tradeoffs for long prompts and large contexts.

Generative AI and foundation model review

For 2026 PMLE preparation, treat generative AI as part of production ML engineering: data grounding, evaluation, safety, latency, cost, and governance matter more than prompt cleverness alone.

NeedReview approach
Summarization or generationUse a foundation model through a managed platform such as Vertex AI when appropriate.
Domain-specific Q&AConsider retrieval-augmented generation using embeddings, vector search, and grounded context.
Semantic searchGenerate embeddings and search by vector similarity.
Safer outputsUse grounding, safety controls, content filtering, prompt constraints, and human review.
Better domain behaviorCompare prompt engineering, RAG, supervised tuning, or other adaptation methods based on data and risk.
EvaluationMeasure relevance, factuality, groundedness, toxicity/safety, latency, and user satisfaction.
Cost controlCache where appropriate, reduce prompt size, choose model size carefully, batch offline jobs when possible.
GovernanceLog prompts/responses carefully, protect sensitive data, and define retention policies.

Generative AI traps

  • Sending confidential data to a model without checking privacy and access requirements.
  • Assuming generated text is factual without grounding or validation.
  • Evaluating only with subjective examples instead of a repeatable test set.
  • Using a large general model when embeddings, search, or a smaller model would solve the problem.
  • Ignoring prompt injection, unsafe content, data leakage, or hallucination risk.
  • Treating RAG as automatic truth rather than a system that needs retrieval quality, chunking strategy, and evaluation.

Vertex AI lifecycle checkpoints

Lifecycle stepWhat to remember for PMLE
Dataset creationValidate schema, labels, splits, and access permissions before training
TrainingChoose AutoML, custom training, BigQuery ML, or foundation model approach based on constraints
Hyperparameter tuningUse when model class is appropriate but performance depends on configuration
ExperimentsTrack parameters, metrics, source version, data version, and artifact location
Model RegistryVersion and promote models through review stages
DeploymentChoose endpoint, batch prediction, or custom serving target based on latency and throughput
Traffic splittingUse for canary, A/B test, or gradual rollout where supported
MonitoringTrack skew, drift, prediction distribution, latency, errors, and business KPIs
RetrainingAutomate with pipeline triggers, validation gates, and rollback plan
Notes and examples

Custom training pattern

Use custom training when you need control over model code, libraries, training loop, hardware, or distributed strategy.

High-yield components:

ComponentPurpose
Training containerReproducible runtime with dependencies
Training service accountReads training data, writes artifacts, logs metrics
Cloud Storage / Artifact RegistryStores packages, containers, model artifacts
Vertex AI custom jobRuns managed training workload
Vertex AI hyperparameter tuningSearches parameter space with managed trials
Vertex AI Experiments / MetadataTracks run lineage and metrics

Minimal command-shape recognition:

gcloud ai custom-jobs create \
  --region=REGION \
  --display-name=JOB_NAME \
  --worker-pool-spec=machine-type=MACHINE_TYPE,replica-count=1,container-image-uri=IMAGE_URI

Do not memorize flags as the main skill. For the exam, understand why a custom job is chosen and which service account, data path, artifact path, and region it uses.

BigQuery ML quick patterns

Use BigQuery ML when the training data is already in BigQuery and SQL-native training/scoring satisfies the requirement.

CREATE OR REPLACE MODEL `project.dataset.churn_model`
OPTIONS(
  model_type = 'BOOSTED_TREE_CLASSIFIER',
  input_label_cols = ['churned']
) AS
SELECT
  * EXCEPT(customer_id, split)
FROM `project.dataset.training_features`
WHERE split = 'TRAIN';
SELECT
  *
FROM ML.EVALUATE(
  MODEL `project.dataset.churn_model`,
  (
    SELECT * EXCEPT(customer_id, split)
    FROM `project.dataset.training_features`
    WHERE split = 'TEST'
  )
);
SELECT
  customer_id,
  predicted_churned,
  predicted_churned_probs
FROM ML.PREDICT(
  MODEL `project.dataset.churn_model`,
  (
    SELECT * EXCEPT(churned, split)
    FROM `project.dataset.scoring_features`
  )
);

BigQuery ML exam clues:

ClueInterpretation
Analysts prefer SQLBigQuery ML is likely
Model training data is warehouse-nativeAvoid unnecessary export
Need batch scoring into tablesBigQuery ML or Vertex AI batch prediction
Need low-latency endpointBigQuery ML alone is usually not the best fit
Need complex custom neural architectureUse Vertex AI custom training instead

MLOps, CI/CD, and reproducibility

What to version

AssetWhy it matters
Source codeReproduce training and serving behavior
Container imageReproduce runtime dependencies
Data snapshot or queryReproduce training set
Feature definitionsPrevent train/serve mismatch
HyperparametersExplain metric differences
Metrics and evaluation reportsCompare candidate models
Model artifactPromote, rollback, and audit
Pipeline definitionRecreate workflow
Service account and IAM changesInvestigate access and security issues
Notes and examples

Deployment patterns

PatternUse whenWatch for
Manual deploymentPrototype onlyNot reproducible or auditable
CI/CD to staging then productionProduction ML serviceAdd evaluation and approval gates
Canary deploymentValidate new model on small traffic shareMonitor error rates and KPIs
A/B testingCompare business impactRequires experiment design and unbiased assignment
Shadow deploymentObserve new model without affecting usersNeeds duplicate inference and log analysis
Blue/greenFast rollback between versionsRequires parallel environment readiness
Batch replacementPeriodic scoring pipelineValidate output schema and downstream consumers

Pipeline validation gates

GateFailure should block
Data schema checkTraining with incompatible data
Data quality thresholdTraining on corrupt or incomplete data
Minimum evaluation metricPromoting weak model
Fairness/slice metric regressionShipping model harmful to a segment
Latency/load testDeploying model that cannot serve traffic
Security scanDeploying vulnerable container or dependency
Explainability or review requirementReleasing opaque high-risk model

Deployment patterns

Prediction needRecommended patternNotes
Low-latency per-request predictionsOnline prediction endpointUse managed Vertex AI endpoints when standard serving is sufficient.
Large scheduled scoring jobsBatch predictionBetter for offline scoring, reports, recommendations, or periodic risk scoring.
Event-driven scoringPub/Sub with Dataflow, Cloud Run, or other processingUseful for streaming use cases and decoupled ingestion.
Embedded app-specific inferenceCloud Run or GKEUse when serving requires custom APIs, routing, or orchestration.
Heavy model with specialized hardwareEndpoint with appropriate acceleratorValidate latency, throughput, cost, and autoscaling behavior.
Edge or disconnected inferenceExported or optimized modelConsider model size, update mechanism, and device constraints.

Deployment decision rules

  • If the model is called synchronously by an application, think online prediction.
  • If millions of records are scored overnight, think batch prediction.
  • If events arrive continuously, think streaming pipeline.
  • If the question emphasizes managed ML lifecycle, think Vertex AI.
  • If the question emphasizes custom application serving, networking, or microservices, consider Cloud Run or GKE.
  • If the question emphasizes rollback safety, choose canary, blue/green, or versioned endpoint deployment.

Security, IAM, and governance

Security controls matrix

ConcernGoogle Cloud control patternPMLE decision point
Least privilegeUse service accounts with minimal rolesAvoid broad Owner/Editor grants
Human accessGrant developers only needed Vertex AI, BigQuery, Storage, and service account permissionsSeparate human identity from runtime identity
Runtime identityDedicated service account for training, pipeline, and predictionDo not run jobs with personal credentials
Sensitive dataBigQuery policy tags, row/column-level controls, DLP-style inspection where appropriateProtect training data and features
Network boundaryPrivate access patterns, VPC controls where requiredAvoid public exposure for sensitive workloads
Encryption controlGoogle-managed encryption by default; CMEK where requiredEnsure services and artifacts support required key usage
SecretsSecret ManagerDo not bake secrets into images, notebooks, or environment files
AuditabilityCloud Audit Logs, Cloud LoggingNeeded for regulated or high-risk ML workflows
Container supply chainArtifact Registry, build scanning, pinned dependenciesReproducible and reviewable deployments
Data exfiltration riskVPC Service Controls where appropriateEspecially relevant for managed service access to sensitive data
Notes and examples

IAM role-shape reminders

PrincipalNeedsAvoid
Data scientistSubmit jobs, read approved datasets, view experimentsBroad project admin
Training job service accountRead training data, write model artifacts, write logsAccess to unrelated production data
Pipeline service accountOrchestrate pipeline components and pass approved runtime accountsAbility to modify all IAM
Prediction serviceServe model and write logs/metricsAccess to raw training data unless required
CI/CD service accountBuild containers, push artifacts, deploy approved modelsPersonal credentials or unrestricted production access
Monitoring operatorRead metrics/logs, acknowledge alertsAbility to alter training data or models unnecessarily

Monitoring and troubleshooting

Production ML monitoring signals

SignalWhat it detects
Prediction latencyServing bottlenecks, oversized model, cold starts, downstream delays
Error rateBad inputs, container failures, dependency issues
Input feature distributionData drift, schema changes, source system changes
Training-serving skewDifferent feature logic or data freshness between train and serve
Prediction distributionCollapsed model, threshold issue, unexpected population shift
Ground-truth performanceReal model quality after labels arrive
Slice metricsDegradation for specific user, region, product, or demographic segment
Resource utilizationUnder/over-provisioning, accelerator bottlenecks
Business KPIWhether ML improvement matters operationally
Notes and examples

Troubleshooting runbook

SymptomLikely causesFirst actions
Custom training job fails immediatelyBad container entrypoint, missing dependency, IAM denial, invalid pathCheck logs, image URI, service account permissions, artifact locations
Training cannot read dataRuntime service account lacks BigQuery/Storage accessGrant least-privilege read to the training service account
Out-of-memory during trainingBatch too large, model too large, inefficient input pipelineReduce batch size, use larger machine, optimize data loading
Training is slowInput bottleneck, no accelerator use, poor sharding, cross-region dataCo-locate resources, optimize input pipeline, profile workload
Great validation, poor productionLeakage, skew, nonrepresentative split, stale featuresRebuild split, compare train/serve features, inspect production distribution
Accuracy high but business impact poorWrong metric, imbalance, bad thresholdOptimize metric aligned to cost and decision process
Endpoint latency highModel size, inefficient preprocessing, no batching, scaling configProfile preprocessing/model, use efficient serving container, tune scaling
Drift alert firesSource distribution changed, upstream bug, seasonalityValidate data source, compare slices, retrain only after quality review
Pipeline not reproducibleUnversioned data/code/image, nondeterministic splitPin versions, use stable split, log parameters and artifacts
Permission error in pipelineWrong runtime account or missing pass-through permissionIdentify executing principal and grant minimal required role

Responsible AI and explainability

TopicExam-ready action
FairnessEvaluate metrics by relevant slices, not only aggregate score
ExplainabilityUse feature attribution/explanations where supported and meaningful
Bias in labelsInspect label source and sampling process
PrivacyMinimize sensitive features and control access to raw data
Human oversightAdd review for high-impact automated decisions
Model cards / documentationRecord intended use, limitations, metrics, training data summary
Safety for generative AIEvaluate harmful content, hallucination, leakage, and prompt injection
MonitoringWatch for drift and quality regressions after deployment
Feedback loopsAvoid model decisions contaminating future labels without controls

Common trap: “The model has high AUC, so it is ready.” PMLE-style answers often require slice evaluation, threshold selection, explainability, security review, and production monitoring before release.

Notes and examples

Responsible AI and explainability

PMLE scenarios may ask for a technically sound model that is also safe, fair, interpretable, and governable.

ConcernPractical response
Bias in training dataAnalyze representativeness, label quality, and slice performance.
Unequal error ratesEvaluate metrics by subgroup; adjust data, thresholds, or model strategy.
Explainability requirementUse interpretable models, feature attribution, example explanations, or documentation.
Human impactAdd human review for high-risk decisions.
TransparencyDocument model purpose, limitations, data sources, and evaluation results.
Monitoring fairnessTrack production performance across relevant slices when labels are available.
Feedback loopsWatch for models that influence future training data, such as recommendations or moderation systems.

Architecture patterns to recognize

Managed tabular prediction

LayerTypical choice
DataBigQuery
TrainingVertex AI AutoML or BigQuery ML
PipelineVertex AI Pipelines
RegistryVertex AI Model Registry
ServingVertex AI endpoint for online; BigQuery ML or batch prediction for offline
MonitoringVertex AI monitoring plus business KPI tracking
Notes and examples

Choose this when the task is standard tabular ML and custom architecture is not required.

Streaming fraud or anomaly scoring

LayerTypical choice
IngestionPub/Sub
Feature computationDataflow streaming
StorageBigQuery for analytics; feature store/low-latency store if needed
ServingVertex AI endpoint or custom low-latency service
MonitoringLatency, error rate, precision/recall after labels arrive

Key traps: class imbalance, delayed labels, duplicate events, threshold tuning, false positive cost.

Batch forecasting

LayerTypical choice
Historical dataBigQuery or Cloud Storage
Feature creationBigQuery SQL, Dataflow, or pipeline component
TrainingBigQuery ML, AutoML, or custom training depending on complexity
ScoringScheduled batch prediction
OutputBigQuery table for downstream planning
ValidationTime-based backtesting

Key trap: random split leaks future data.

Document-grounded generative AI

LayerTypical choice
Document ingestionControlled storage and indexing pipeline
RetrievalManaged search/vector retrieval pattern
GenerationVertex AI foundation model
ControlsGrounding, citations, safety settings, prompt injection defenses
EvaluationGroundedness, factuality, task completion, human review

Key trap: fine-tuning a model to memorize private documents when retrieval is the safer, fresher design.

Common PMLE traps

TrapBetter answer
Choose custom Kubernetes for every ML workloadPrefer Vertex AI managed services unless requirements demand custom orchestration
Optimize accuracy on imbalanced dataUse precision, recall, F1, PR AUC, threshold tuning
Randomly split time-series dataUse time-based validation and backtesting
Ignore training-serving skewReuse transformation logic or centralize feature definitions
Train with personal credentialsUse dedicated service accounts
Store secrets in notebooks or containersUse Secret Manager
Deploy model without monitoringAdd latency, errors, drift/skew, and quality monitoring
Fine-tune LLM to add changing factsUse RAG/grounding for factual enterprise knowledge
Export BigQuery data unnecessarilyUse BigQuery ML or native integrations when suitable
Use online prediction for offline bulk scoringUse batch prediction or warehouse scoring
Compare models on different data splitsUse consistent test data and logged experiments
Promote model based only on aggregate metricCheck slices, business cost, fairness, and operational constraints
Notes and examples

Common PMLE scenario traps

TrapBetter reasoning
Choosing the newest or most complex servicePrefer the simplest managed option that meets requirements.
Optimizing accuracy for imbalanced dataUse metrics aligned to positive-class and business costs.
Randomly splitting time-series dataUse time-based validation to avoid future leakage.
Training and serving with separate preprocessing logicShare transformations and version preprocessing artifacts.
Deploying after validation onlyAdd monitoring, rollback, and production guardrails.
Using batch prediction for low-latency app callsUse online prediction when synchronous latency matters.
Using online prediction for massive scheduled scoringUse batch prediction to reduce operational overhead.
Scaling compute before fixing data pipeline bottlenecksCheck input pipeline, preprocessing, and storage throughput.
Retraining automatically on bad dataValidate data before training and gate deployment on evaluation.
Granting broad permissions to simplify setupUse least privilege and service-account separation.
Ignoring labels that arrive lateDesign delayed ground-truth evaluation and monitoring.
Assuming offline improvement guarantees business improvementUse canary, A/B testing, or business KPI validation.
Not versioning datasetsReproducibility requires dataset, code, config, and artifact versions.
Using foundation models without safety evaluationAdd groundedness, safety, privacy, and human-risk checks.

Last-minute checklist

Before answering a PMLE scenario question, identify:

  • Task: classification, regression, forecasting, ranking, generation, clustering, anomaly detection.
  • Data location: BigQuery, Cloud Storage, streaming, external source.
  • Latency: online request, near-real-time stream, scheduled batch.
  • Control level: AutoML, BigQuery ML, custom training, foundation model, pretrained API.
  • Metric: aligned to business cost and class balance.
  • Split: leakage-resistant and time-aware if needed.
  • Pipeline: reproducible, versioned, and automated.
  • Security: service accounts, least privilege, sensitive data controls.
  • Deployment: endpoint, batch job, warehouse scoring, or custom serving.
  • Monitoring: drift, skew, latency, errors, ground-truth performance.
  • Rollback: model versioning, canary/shadow/blue-green where appropriate.

PMLE Cheat Sheet focus

This Cheat Sheet is for candidates preparing for Google Cloud’s Professional Machine Learning Engineer (PMLE) exam. It is IT Mastery review support, not affiliated with Google Cloud, and is designed to help you quickly reinforce high-yield concepts before using topic drills, mock exams, and detailed explanations.

For PMLE, do not study machine learning as isolated algorithms only. The exam is usually most challenging when it asks you to choose a practical Google Cloud design that balances model quality, reliability, security, cost, monitoring, and operational maintainability.

Use this page to review:

  • How to frame ML problems and choose evaluation metrics.
  • When to use Vertex AI, BigQuery ML, Dataflow, Dataproc, Cloud Storage, Pub/Sub, GKE, Cloud Run, and related Google Cloud services.
  • How to prepare data, avoid leakage, and reduce training-serving skew.
  • How to deploy, monitor, retrain, and govern models in production.
  • How to reason through scenario questions without memorizing product trivia.

Problem framing and metrics

PMLE scenarios often test whether you choose the right objective before choosing tools. A technically sophisticated model can still be wrong if it optimizes the wrong metric.

Problem typeUseful metricsCommon traps
Binary classificationPrecision, recall, F1, ROC AUC, PR AUC, log lossAccuracy can be misleading with class imbalance.
Multiclass classificationMacro/micro F1, top-k accuracy, confusion matrixOverall accuracy can hide poor minority-class performance.
RegressionMAE, RMSE, RMSLE, R-squaredRMSE over-penalizes large errors; MAE may be better when robustness matters.
Ranking/recommendationNDCG, MAP, MRR, CTR, conversion rateOffline ranking metrics may not match user behavior in production.
ForecastingMAE, RMSE, MAPE, WAPE, MASERandom splits can leak future information.
Anomaly detectionPrecision, recall, PR AUC, false positive rateRare events make accuracy nearly useless.
ClusteringSilhouette score, Davies-Bouldin, business validationUnsupervised metrics do not guarantee useful segments.
Generative AI outputGroundedness, factuality, safety, relevance, human preferenceBLEU-like text metrics may not capture business risk or factual correctness.
Notes and examples

Key classification formulas:

\[ \text{Precision} = \frac{\text{True Positives}}{\text{True Positives} + \text{False Positives}} \]\[ \text{Recall} = \frac{\text{True Positives}}{\text{True Positives} + \text{False Negatives}} \]\[ \text{F1} = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} \]

Metric decision rules

If the scenario says…Prioritize…
“False positives are expensive”Precision or specificity.
“Missing a positive case is dangerous”Recall or sensitivity.
“Classes are highly imbalanced”PR AUC, F1, class-weighted metrics, stratified evaluation.
“Predicted probabilities are used for decisions”Calibration and log loss, not just class labels.
“Large errors are especially bad”RMSE or custom loss.
“Outliers should not dominate”MAE or robust loss.
“Business cost differs by error type”Custom cost function or threshold optimization.
“Model must rank best candidates first”Ranking metrics such as NDCG or MAP.

Threshold trap

Many candidates assume a classification threshold of 0.5. In production, the threshold should usually be selected based on business cost, precision-recall tradeoff, capacity constraints, or risk tolerance. Training a model and choosing an operating threshold are separate decisions.

Model selection: managed, custom, or foundation model?

PMLE questions often include clues about team skills, time constraints, explainability, customization, data volume, latency, and governance.

ApproachUse whenAvoid when
Pretrained Google Cloud APIsStandard tasks such as vision, speech, translation, document, or language extraction fit the use case.You need deep customization, domain-specific labels, or strict control over model internals.
Vertex AI AutoMLYou need strong baseline performance with limited ML engineering effort.You require custom architecture, unusual loss functions, or specialized training loops.
BigQuery MLData is already in BigQuery and the model can be built using supported SQL-based workflows.The workload requires complex custom deep learning or custom serving logic.
Custom training on Vertex AIYou need custom code, frameworks, tuning, distributed training, or specialized containers.A managed AutoML or pretrained option satisfies requirements more simply.
Imported model on Vertex AIYou already have a trained model and want managed deployment/serving.The model needs substantial retraining or incompatible serving dependencies.
GKE or Cloud Run custom servingYou need custom inference orchestration, special networking, or app-specific serving behavior.A standard managed Vertex AI endpoint is sufficient.
Foundation model through Vertex AIUse cases involve generation, summarization, chat, extraction, embeddings, or semantic search.Deterministic, low-risk, fully explainable traditional ML is required and generation adds unnecessary risk.
Notes and examples

AutoML versus custom training

Choose AutoML when the exam scenario emphasizes:

  • Fast development.
  • Limited ML expertise.
  • Standard tabular, image, text, or video use cases.
  • Managed training and tuning.
  • A strong baseline without custom architecture.

Choose custom training when it emphasizes:

  • Custom loss functions or metrics.
  • Specialized model architectures.
  • Complex preprocessing or training loops.
  • Distributed training.
  • Framework-specific requirements.
  • Full control over dependencies and containers.

Training, tuning, and optimization

Overfitting versus underfitting

SymptomLikely issuePractical fix
High training performance, poor validation performanceOverfittingMore data, regularization, dropout, simpler model, early stopping, augmentation.
Poor training and validation performanceUnderfittingMore expressive model, better features, longer training, lower regularization.
Validation performance unstableSmall validation set or noisy labelsBetter split, cross-validation, label review, more data.
Great offline metrics, poor production resultsSkew, leakage, drift, or wrong metricValidate feature availability, monitor production, reassess metric.
Model improves but latency is too highServing inefficiencyOptimize model, use batch prediction, quantization, distillation, accelerators, or simpler architecture.
Notes and examples

Hyperparameter tuning review

High-yield hyperparameters:

  • Learning rate.
  • Batch size.
  • Number of layers or trees.
  • Regularization strength.
  • Dropout rate.
  • Maximum tree depth.
  • Embedding dimension.
  • Optimizer choice.
  • Early stopping patience.

Common traps:

  • Tuning on the test set.
  • Reporting the best validation score as final test performance.
  • Ignoring cost and time of large tuning jobs.
  • Changing data preprocessing during tuning without versioning it.
  • Optimizing a proxy metric that does not match the business objective.

Distributed training and accelerators

RequirementReview answer
Large neural network trainingConsider GPU or TPU acceleration depending on framework and workload fit.
Training data too large for one workerUse distributed training or data-parallel approaches.
CPU-bound preprocessingOptimize input pipeline; accelerators do not fix slow data loading.
Low GPU utilizationCheck batch size, input pipeline, data transfer, and model size.
Cost concernUse managed jobs, right-sized machines, early stopping, preemptible/spot-style strategies where appropriate, and avoid idle accelerators.

Evaluation and validation

A PMLE-ready evaluation plan includes more than a single score.

Evaluation layerWhat to check
Holdout test performanceFinal unbiased estimate after tuning.
Cross-validationUseful when data is limited or variance is high.
Slice performancePerformance across regions, devices, languages, demographic groups, product categories, or customer segments.
CalibrationWhether predicted probabilities match observed frequencies.
FairnessWhether errors or outcomes are disproportionately harmful across groups.
RobustnessSensitivity to noise, missing values, outliers, prompt variation, or distribution shift.
ExplainabilityFeature attribution, example-based explanations, model cards, stakeholder interpretability.
Latency and throughputWhether model quality is achievable under serving constraints.
CostTraining cost, prediction cost, storage, orchestration, and monitoring overhead.
Notes and examples

Offline versus online evaluation

MethodPurposeTrap
Offline validationCompare models before deployment.May not predict user behavior or business impact.
Shadow deploymentSend production traffic to new model without affecting users.Does not prove user response changes because outputs are not acted on.
Canary deploymentServe small traffic percentage to new model.Needs rollback and monitoring.
A/B testMeasure causal business impact.Requires careful experiment design, sample size, and guardrail metrics.
Blue/green deploymentSwitch between full environments.Useful for rollback but not always enough for model behavior validation.

MLOps, reproducibility, and pipelines

MLOps questions reward operational discipline.

MLOps needWhat good looks like
Reproducible trainingVersion data, code, dependencies, parameters, containers, and model artifacts.
Automated workflowUse pipelines for data validation, training, evaluation, approval, deployment, and monitoring.
Model governanceRegister models, track lineage, document metrics, require approvals where needed.
Safe deploymentPromote models through environments, use CI/CD, validate before serving traffic.
RollbackKeep previous model versions and serving configs available.
AuditabilityLog who changed data, code, parameters, model versions, and deployments.
Continuous trainingTrigger retraining based on schedule, new data, drift, or performance degradation.
Experiment trackingCompare runs consistently using parameters, metrics, artifacts, and dataset versions.
Notes and examples

Pipeline anti-patterns

Avoid answers that:

  • Manually run notebooks for production training.
  • Deploy models without validation gates.
  • Overwrite model artifacts without versioning.
  • Use broad owner permissions for pipeline service accounts.
  • Store secrets in source code or container images.
  • Retrain automatically without checking model quality before deployment.
  • Ignore rollback when changing models used by production systems.

Monitoring and production reliability

Production ML monitoring includes software reliability and model behavior.

MonitorWhy it matters
Request countDetect traffic spikes or drops.
Latency percentilesp95/p99 latency often matters more than average latency.
Error rateDetect serving failures, dependency failures, or malformed requests.
Resource utilizationIdentify CPU, memory, GPU, or autoscaling issues.
Input schemaCatch missing fields, type changes, and invalid ranges.
Feature distributionDetect skew or data drift.
Prediction distributionDetect sudden output changes.
Ground-truth performanceValidate actual accuracy when labels become available.
Business KPIsConfirm model improvements translate into business value.
Fairness slicesDetect degradation for specific subgroups.
Notes and examples

Drift versus skew versus concept drift

TermMeaningExampleResponse
Training-serving skewTraining and serving data or transformations differ.Feature computed in batch training but not available online.Fix pipeline consistency and shared transformations.
Data driftInput distribution changes over time.Users from a new region create different feature values.Monitor distributions, retrain or adapt features.
Concept driftRelationship between features and label changes.Fraud patterns change after attackers adapt.Retrain with recent labels, update strategy, monitor performance.

Retraining triggers

Retrain when:

  • Ground-truth performance falls below an accepted threshold.
  • Data drift is significant and affects model quality.
  • New labeled data materially improves coverage.
  • Product behavior or business rules change.
  • A fairness, safety, or compliance issue appears.
  • A better model passes validation and operational checks.

Do not retrain blindly if the root cause is a broken upstream pipeline, serving bug, label delay, or schema change.

Security, privacy, and access control

PMLE candidates should connect ML architecture to Google Cloud security fundamentals.

AreaReview focus
IAMGrant least privilege to users, service accounts, pipelines, and serving systems.
Service accountsUse workload-specific identities instead of broad shared accounts.
SecretsStore in Secret Manager; do not hard-code in notebooks, images, or repositories.
EncryptionUse Google Cloud encryption defaults and customer-managed keys where requirements specify.
Network controlsUse private connectivity and service perimeters when sensitive data boundaries matter.
Data minimizationUse only necessary fields; remove or mask sensitive attributes when not needed.
PII handlingDetect, classify, de-identify, tokenize, or redact sensitive data where appropriate.
Audit loggingTrack access to data, artifacts, pipelines, and deployments.
Artifact securityStore images and packages in managed registries with scanning and access control.
Separation of dutiesKeep development, approval, and production deployment roles distinct when governance requires it.
Notes and examples

Security traps

  • Giving a training pipeline broad project owner permissions.
  • Exporting sensitive training data to unmanaged locations.
  • Putting API keys in notebooks or container images.
  • Allowing production models to read more data than required.
  • Ignoring audit requirements for model artifacts and data lineage.
  • Using public endpoints when private access is required by the scenario.

Quick symptom-to-fix table

Symptom in questionLikely causeStrong answer direction
Validation score high, production score poorLeakage, skew, or driftCompare training and serving data; monitor features; fix pipeline.
Model misses rare positive casesImbalanced data or wrong thresholdOptimize recall/PR AUC; resampling, class weights, threshold tuning.
Too many false alertsPrecision problemAdjust threshold, improve features, use cost-sensitive evaluation.
Users complain about slow predictionsServing latencyOptimize model, use accelerators, autoscaling, caching, or batch prediction.
Training job slow with idle GPUInput bottleneckImprove data loading, preprocessing, batching, and storage throughput.
Model quality differs by region/languageSlice performance issueEvaluate by subgroup; improve data coverage and monitoring.
Pipeline sometimes deploys bad modelsMissing validation gateAdd automated evaluation and approval criteria.
Model degrades after product changeConcept or data driftMonitor, retrain, update features, validate new behavior.
Sensitive data appears in logsPrivacy control failureRedact, minimize logging, protect access, review retention.
Generated answers are plausible but wrongHallucination or weak groundingUse RAG, citations, evaluation, safety checks, human review.

Final review checklist

Before moving to PMLE question-bank practice, make sure you can answer these quickly:

  • Can you map a business goal to the right ML task and metric?
  • Can you explain why accuracy may be the wrong metric?
  • Can you choose between AutoML, BigQuery ML, custom Vertex AI training, pretrained APIs, and foundation models?
  • Can you identify feature leakage and training-serving skew?
  • Can you choose the correct split strategy for time series, users, groups, or imbalanced classes?
  • Can you design a reproducible training pipeline with versioned artifacts?
  • Can you select online, batch, or streaming prediction based on latency and volume?
  • Can you describe safe rollout, rollback, monitoring, and retraining?
  • Can you apply IAM least privilege to ML pipelines and model serving?
  • Can you address privacy, explainability, fairness, and responsible AI requirements?
  • Can you evaluate generative AI systems for groundedness, safety, relevance, cost, and latency?

Practice plan after this Cheat Sheet

Use IT Mastery practice to convert this review into exam readiness:

  1. Start with topic drills on weak areas: metrics, data leakage, Vertex AI services, deployment, monitoring, security, and responsible AI.
  2. Review every missed question with detailed explanations, especially why the wrong answers are tempting.
  3. Move to mixed original practice questions once individual topics feel stable.
  4. Use full mock exams to practice scenario triage, time management, and eliminating overbuilt solutions.
  5. Revisit this Cheat Sheet after each mock exam and update your personal trap list.

Next step: begin targeted PMLE question bank practice with original practice questions, then use detailed explanations to close gaps before attempting full-length mock exams.

Put the review into practice